Total Recall: Agent Memory and Harness Engineering — Ignacio Martinez, Oracle

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Total Recall: Agent Memory and Harness Engineering — Ignacio Martinez, Oracle · Channel: AI Engineer · 1:00:20 · Watch on YouTube

Topics: AI agents, Data, databases and embeddings, Prompt and context engineering, RAG, search and information retrieval

Summary

Ignacio Martinez introduces agent harnesses as the systems around a language model that add memory, tools, perception, and structure to make outputs more reliable and repeatable. He surveys the agent stack, then explains harness components including storage, memory engineering, semantic context, agent loops, and context engineering. He argues for combining files with databases and describes approaches to short-term, long-term, and shared memory, as well as skill promotion and continual learning. The session concludes with a workshop demonstration and discussion of tool retrieval, agent-loop limits, and model routing.

Key points

An AI agent can be understood as a language model plus a harness that supplies memory, tools, and environmental inputs.
Harness engineering aims to make the non-deterministic outputs of models more reliable and repeatable.
The agent stack includes the application, data, model, infrastructure, and compute; Martinez identifies data as the layer developers can control most.
The seven harness components discussed include storage, memory engineering, semantic context, agent loops, and context engineering.
Files are easy for models and operating systems to work with, while databases provide capabilities such as transactional consistency, backups, and vector search; a hybrid approach can combine their benefits.
Memory can be short-term, long-term, or shared, and may include episodic memories and reusable procedural workflows.
Large context windows can suffer from context rot, so context should stay focused and relevant to the task.
The semantic layer captures organizational knowledge and other assumptions that users might otherwise leave unsaid.
Skill and workflow promotion can turn successful work into reusable, improvable instructions.
The workshop demonstrates a working harness and discusses tool retrieval, bounded tool calls, and routing tasks to different models.

Chapters

Start	Chapter	Summary
00:12	Workshop setup and plan	Martinez explains how attendees can access the repository and says the session will begin with an introduction before moving into the workshop.
03:22	Introduction to agent harnesses	He outlines the session goals and introduces the agent stack, emphasizing the data layer as a key area of control.
08:02	AI application shapes and the agent definition	He describes gateways and MCP, distinguishes kinds of AI applications, and defines an agent as a model plus a harness.
13:48	The seven components of an agent harness	Martinez previews harness components, including storage, memory engineering, semantic context, agent loops, and context engineering.
15:24	Files, databases, and hybrid storage	He compares file-based and database storage and advocates combining them to support different memory needs.
21:34	Embeddings, retrieval, and data management	He explains embedding search and reranking, then discusses the engineering overhead of managing multiple kinds of AI data.
25:45	Agent memory and context rot	Martinez covers memory types and explains why simply expanding the context window does not replace memory engineering.
30:27	Memory engineering and context cards	He presents memory engineering and the Oracle Agent Memory Package as ways to organize and assemble relevant conversational context.
34:04	The semantic layer and agent loops	He explains how a semantic layer captures implicit organizational knowledge, then introduces the agent loop and context patterns.
38:13	Continual learning and skill promotion	Martinez describes ways to improve agent behavior over time by promoting successful workflows and skills.
41:50	Starting the hands-on workshop	He directs attendees to create a GitHub Codespace and begins setting up the workshop materials.
44:30	App book and workshop environment	He introduces the app book and explains how the workshop will use Oracle’s managed model service and a student notebook.
49:12	Exploring the harness demonstration	Martinez demonstrates the app’s context visualization, selected tools, traces, and fault-tolerant agent behavior.
51:19	Tool retrieval and model routing	He answers questions about finding tools at scale, handling tool descriptions, setting limits on tool calls, and routing tasks to models.
58:15	Notebook guidance and closing	He explains how to run the student notebook, points attendees to its documentation, and invites them to continue the discussion.

Quotes

An AI agent is essentially a model large language model plus a harness. — 10:08

The goal of harness engineering is to create reliable and predictable outputs over and over. — 11:11

You don’t actually need to choose one or the other. You can actually use both of them. — 17:28

You want to keep the context window as small as possible to avoid context rot. — 29:55

Concepts and entities

Ignacio Martinez, Oracle, Andrew Ng, agent harness, agent memory, Oracle Agent Memory Package (OAMP), Oracle DBFS, LangChain OracleDB, OCI Generative AI service, MCP, context rot, toolbox pattern, HNSW

Full transcript (original language)

00:12 Hello, hello. Perfect. Yep. So, perfect. Yeah, this this volume is perfect. It’s just a website that I registered the domain on Saturday just to do the registration of the workshop. But, for those of you who are going to follow along with me, just know that after you complete that that process in the in the workshop, you will get an invitation to the repository. And just view the invitation, accept it,

00:42 and then you’ll be able to run the workshop. We’ll be running the workshop on GitHub code spaces. So, if you want to start that up, it takes like 5 minutes. Um just for reference, I’m going to use the first 30 minutes to give you an introduction to Agent Harness and lots of other concepts, and then in the next hour and a half we’re going to go actually go through uh the workshop together. Sound good? Okay, perfect. And thank you, by the way, for being here because I know it’s

01:14 Monday 9:00 a.m. So, I commend you all for for being here. >> Uh just 2 minutes before we begin, so I will shut up.

03:22 >> All right. Well, let’s let’s begin, guys. So, thank you for being on this uh at this time here with me. Um I know it’s 9:00 a.m. like I said, so hopefully I can and and we me and my team can make your time worthwhile. At the end of this session, what I would like you to leave with is some knowledge on agent memory and agent harnesses, how you can build your own agent harness, which is nowadays one of the hot topics on AI, I would say.

03:55 Um Lots of people are talking about models constantly, but the thing is that models, like language models, they are the frozen part of the reasoning, right? We just have to accept what we are given. And during the past few weeks, you will if you’re following the news, you will have seen that this is this has never been more true than now. So, the harness is what we’re going to talk about. For those of you who weren’t here, uh you can scan this QR code or go to

04:26 that website workshopwaitingroom.com and just register. You will get an invitation to a GitHub repo. And on this GitHub, you will be able to create a GitHub code space where we will run the workshop. And you’ll have everything set up for you. So, just a little introduction on on who I am. I’ve been working for Oracle for 7 years. I’ve been a developer advocate for about four of them.

04:56 And you can find, you know, my talks and I’m very active on GitHub as well. So, if you’re a GitHub user, just check out my GitHub profile if you like. Uh what is the highlight of my career so far? I launched a course with Andrew Ng on agent memory. So, if you’re interested on the memory components of what we’re going to what we are going to discuss today, you can just check out that course if you’d like. Um these are the things that we are going

05:27 to talk through today, right? First of all, we’re going to do a little introduction on what the agent stack is, and then we’re going to zoom into the data layer where the memory lives. Then we’re going to explore a little bit about the shapes, the different shapes that AI applications take nowadays, uh what an agent is uh followed by the seven different layers that make up an agent harness. So, we if you just follow these seven different structures, you will be able

05:58 to create an agent harness, a minimal agent harness that you can connect any model to. Then, we’re going to finally talk a little bit about continual learning and how as as Oracle we are uniquely positioned to help you achieve and develop agent harnesses and and AI applications. So, the agent stack. And by the way, [clears throat] if you have any questions, just raise your hand. I’m very happy to to take questions, as well. So, the agent stack.

06:30 The agent stack, like every agent a every AI agent, sits on on these five layers. You either you have an application, which is the product surface, right? What we interact with as users. We have data. And that has lots of components. You have memory, you have knowledge, you have retrieval, encoding, search. You have the model itself, the reasoning large language model that is behind everything. Uh infrastructure, which is the orchestration. What model do we serve

07:01 depending on the reasoning effort that we need and things like that. And then we have compute, which is the cloud or GPUs and also the the database engine. And the application, the model, the infrastructure, and the compute, these four layers except for the data, they are uh increasing increasingly commoditized. What do I mean by that? They are trying to take away the complexity from these layers out of out of our

07:31 domain, right? So, the thing that we have the most control over when working with AI applications is actually the data. And that’s the part that we’re going going in because the agent harness excels and you know, lives very closely with uh with the data layer. So, let’s focus on the on the data layer. The data layer is where an agent harness appears. And it has, you know, many components. The first component is the gateway and

08:02 MCP. It’s interesting because that’s the one that connects an agent harness to data and tools. So, kind of think of of a large language model like a isolated thing that wouldn’t be able to work at all if it didn’t have access to things like data, right? So, you’ll you you then have like the memory layer, the semantic layer, retrieval layer, context layer, and tools and

08:33 skills that are built on top of the gateway and MCP layer. Let’s say for instance that you have a I don’t know, an application on your computer and your large language model doesn’t have access to it. For instance, Outlook. Right? So, if you might want to have your large language model connect to that, you can create an MCP, you specify some functions, and then the LLM all of a sudden is able to communicate with this program. So, gateway and MCP is like the layer that connects a large language model to

09:05 the outside world or de facto to our computer or wherever you’re working. And AI applications today, they take up four different um four different shapes, right? You have LLM chatbots, which are very passive. Uh they just respond when you ask a question. You also have rag applications, which are semi-passive because they have to do some kind of processing in the background, but then you also have a passive nature of it. And then

09:36 you have the more active components of AI applications, which are what we kind of use every day. Like cloud code, etc., etc., which are a combination of LLM-driven workflows that provide automation and AI agents that provide autonomy. And we’ll go we’ll we’ll explain later what I what I mean by that. But first, I want you to have a very clear definition of what an AI agent is. So, to me, an AI agent is essentially a model

10:08 large language model plus a harness. The model itself will be the reasoning and everything else will be the harness. Um So, this definition is something like this. An autonomous entity whose cognitive functions are powered by a large language model for reasoning. They are augmented by a database or files for memory. They are extended through tools for actions and grounded in inputs that let

10:39 it perceive uh its environment. So, an agent is a a model plus the harness. So, if the agent is the model plus the harness, in this diagram, right? We have reasoning and the reasoning is the part that we don’t control. It’s the part that we are that’s that’s heavily subsidized, the part that we rent. If uh if you’re like me, you’re subscribed to every imaginable subscription on Earth. And that’s the

11:11 thing that we do not control. Right? We have no control over what we are offered. And then we have memory, tools, and perception that actually we get some customizab- Um So, the goal of harness engineering is to create reliable and predictable outputs over and over. Whereas a reasoning model is very non-deterministic. You might give it the same the same input and it

11:42 might produce different outputs every time, right? So, I said that, you know, in AI applications the most typical ones nowadays cloud code, codex uh and any other type that you can think of is a combination of automation plus autonomy. Why? Well, because automation provides reliability to a system and autonomy gives flexibility. And this is a combination that is very

12:12 convenient when we are developing ourselves. By the way, raise of hands, who is working as an AI engineer or as an AI developer? Oh my god, okay. Good. So, you must all have used one of these systems, right? So, all of them they have this commonality, which is they have autonomy and they have flexibility. Um Yeah, so the idea is that these systems, right? They are built

12:44 on top of an a proprietary agent harness. And an agent harness is nothing more than everything that we spoken about an AI agent. All the things that you need to do around that to enable it to produce reliable and repeatable outcomes, right? So, the model itself non-deterministic. Same input, different outputs every time. But the harness, what we want to do with the harness is to turn this non-deterministic nature of a large

13:14 language model and be able to produce reliable and repeatable results. So, the reasoning, which is the part that we do not control, we’re not going to focus. Actually, the harnesses are built on top of models that are kind of swappable. You just need a like a common interface like an OpenAI protocol or the Anthropic API specification, right? All these things make it so that the model part is swappable and the harness is what we’re going to focus on today.

13:48 So, seven things that make up an agent harness. And as I said, we’re not going to touch on the model layer because we have no control over it. But let’s go a little bit more in detail into each of these, right? You need a storage layer on your agent harness that essentially determines where the data is going to live, where the memory physically lives. Uh and we’ll see about this dilemma that has been going on about the last 6

14:20 months about files versus databases. And why I think a hybrid uh combination of both is actually the best part. Then you also have memory engineering components which are all the uh encoding, the search, and the retrieval components of it. And also the semantic layer which is kind of the the the hidden things that happen or the hidden vocabulary that we assume that a large language model knows that is kind of proprietary to our

14:51 companies or our knowledge. What we don’t say to the LLM. Kind of that’s the semantic layer. We’ll lightly touch on agent loops and what an agent loop is and how to implement a very very minimalistic agent loop. And finally, go about some context engineering techniques that you know, keep the window as salient as possible, the context window as salient as possible. You want to minimize the context window as much as possible so that that the task that you’re solving stays relevant.

15:24 So, until here we have done an introduction to what an AI agent is, right? Its use cases. And now we’re going to dive deeper into an agent harness and each one of the individual components. So, the model layer, right? The frozen reasoning core. I say it’s frozen because typically the weights of a model don’t change. Uh and I say typically because uh we are actually I’m actually in the process with Cassius sitting right there. We’re going to

15:55 record a new course with Andrew Ng on continual learning for AI agents. So, if you’re interested, just check that out in a couple couple weeks. But the thing is that 99.9% of time you will have a model and the weights of the model will never change. Uh unless you have millions of dollars or, you know, a lot of time or GPUs, it’s very hard to change the weights of a model. So, there are other ways in which you can affect the reasoning without actually changing the weights of the

16:26 model. But this is motivation for for the continual learning part that we will see in the workshop. Where does the memory live, right? The files versus databases dilemma that we’ve been having uh since January kind of. Some people are very maximalists of files and some of us are Oh, I will not include myself. But some of Some people are also maximalists of the database. Um and you know, both things are right.

16:56 Files have convenient things, convenient characteristics, and also the databases. So, the files, right? They are very easy to like they are very they match the models’ instincts. They are very easy to create. They are very easy to insert and append data into files, right? It’s very It has a very unstructured nature to it. And databases on the other hand, or the conception that people have, is that they have a very structured way, right? But, that’s when you’re

17:28 thinking about SQL, like SQL. And the things that you don’t actually need to ch- like choose one or the other. You can actually use both of them. Like files are attractive because the models speak them, and they kind of work very very easily with operating systems. They follow POSIX semantics, so they are compatible on Debian, Ubuntu, any other operating system that you might want. And

17:58 they also have some disadvantages. For instance, they don’t have transactional consistency. So, this is one of the problems that I wanted to talk about. If you are a degenerate like me, and you are working, I don’t know, like eight, 16, 32 agents at a time, uh the problem with this is that files cannot be modified and inserted and modified at the same time. So, what is the solution nowadays to not having transactional consistency

18:30 and working with files? Any suggestions? Work trees, exactly. So, agents, what they when they want to modify a file, but another agent is working on this thing, they just created a different work tree, right? They will do all the progress in the work tree, and then after the implementation is done, they will merge to master, or the merge to main, sorry. So, this is the way that is a workaround against

19:00 not having transactional consistency, right? And you also have other characteristics, like for instance, hybrid search. Um this is not available on files, but is very very easily achieved on databases. Um you don’t have backups either. So, if your operating system gets corrupted or something, you will just lose everything. And these are things that the database fixed 35 40 years ago, and people have kind of forgotten about that. So, what I

19:30 want to do is to give you kind of the best of each of the of the implementations, right? You can get the benefits of files and the benefits of databases in the same place. And we’ll see why. But these are some of the advantages, right? You have ACID consistency. So, atomic operations, consistent operations, isolated, and durable. You have high availability. You might replicate the database and let it be, you know, in three different places in

20:01 the world with a replication factor. You also have vector search, which is very easy. In files, you just have to do like regular expression matching or a derivative of that. And you know, lots of other lots of other things. So, what we will do on the on the actual workshop is we’re going to run an actual example of trying to modify a file with three different agents that will be working on the same file and trying to

20:31 update a counter on this file. And let’s see who’s faster. I know the answer, of course, but you will you’ll get to know it later. But, what I want to introduce to you is that we have a thing called the Oracle DBFS, or the database file system, where you can store files inside the database on a file system. And that gives you, you know, lots of the advantages that we said. You will get files with ACID transac- transactional consistency. You

21:01 will get vector search, relations, security, high availability, etc. Right? And my suggestion is that since a hybrid system works best, um we can have things like, for instance, uh short-term memory, right? That lives in files. And when something needs to be promoted into a long-term memory, for instance, user preferences, things like this, they can go into a more structured space like a database. And this is what we’ll do in the workshop.

21:34 For the encoding the search and the retrieval, which was another of the components in the agent harness, we also have uh LangChain integration that I want to mention called LangChain OracleDB that makes it very easy to insert into vector stores, search in the vector store, and retrieve from the vector store. So, also we have another thing called in-database embeddings. Have you ever heard about in-database embeddings? Yes? Okay. So, in-database embeddings is very

22:06 convenient, especially for enterprise customers, because you will get the embedding model inside the database, so that when you’re doing embeddings, you don’t have to call a third-party service. And that’s very convenient for isolate like data retention and data security purposes. So, this is what a an an embedding searching and reranking model would look like from a chatbot interface, for instance, right? You have a lot of documents, then you kind of use a

22:37 bi-encoder. So, a bi-encoder is essentially an embedding model. You will create enco- embeddings out of these documents. You will split them and create vectors, put it into a vector store, right? And then you will get a user prompt, a user query, a question. You can also create an embedding out of that and then compare it to what you had previously on your vector store. And this is how you get the most relevant vectors in in an answer. Then you will run an

23:07 across encoder, which is a re-ranker, and take a look at the question plus the result, and that’s how questions are answered kind of in in in rag applications, right? So, these things we are also going to to touch briefly on the workshop. And the thing is that, you know, imagine this rag application from the beginning. From the documents, you create you do lots of things, right? You do tokenization, then you create the embeddings. Um you have to do things

23:40 like deduplicating the data, normalization, uh personal like redacting personally identifiable information, uh lots of things, right? And then on your on your store, you have textual data from the documents. You have metadata, which is typically stored in JSON. You have vectors, which are represented as as dense embeddings of 32 uh bits. Like you have so many types of data that you need to work on that typically what people have is, for instance, I don’t

24:11 know. I will not name names, so I don’t get in trouble, but you know, you might need like different databases for each one of these, right? You you get what I mean? So, there is a very high nowadays, very high data synchronization logic overhead for AI engineers or even for agents, right? So, lots of maintenance required and and and lots of engineering effort on it. And this is something that we want to avoid. So, what I want you to do

24:42 today is just to try us out as Oracle. Uh try our database. We have support for every type of data imaginable that you can think of. Um we are called the converged database. We are the only converged database in the market that we support JSON, we support relational, we support uh spatial, graph, JSON, anything that you can think of, you’ll be able to to create with us. And you’ll get like one database engine, uh one query interface, and uh

25:14 one single development stack. Everything will be in the same database. So, also for data security purposes, you just have to you know, secure and save all your data in one place. So, a single attack vector is what you need to worry about. You don’t need to worry about updating five different database. You can just worry on securing one database. So, we can be the engine of your AI applications, not just a single step,

25:45 which is what people think of when they’re working with databases, right? Um Yeah. So, agent memory. Agent memory is essentially a description of all the mechanisms and the systems that allow an agent to retain, reuse, refine, and recall information. We want to reuse the data and refine it in the process. But we want to reuse the data so that the next time that an agent or

26:18 us as engineers we are working on a problem that took us 3 hours, the next time that we observe this problem, the problem becomes easier either for AI agents or for us. And agent memory has lots of lots of components, right? We have short-term memory, we have long-term memory, and then we have shared memory, which is something that’s relatively new. Um and this shared memory is kind of what happens when a sub agent is communicating with it it’s communicating with its with its parent, for instance,

26:49 or two agents are trying to collaborate on solving one specific problem uh together. And then what I want you to to see is that depending on the type of memory on or or the type of thing that we want to store, it will be in one place or the other, right? So, short-term memory is kind of ephemeral, it’s very short-lived, and it’s very useful for things that are happening right now. For instance, the to-do list on a coding agent, right? It’s happening right now,

27:20 but you actually don’t want to save that uh you know in long-term. But then there are things like, for instance, episodic memory, things that previous conversations that you’ve had. That’s very useful to have, for instance. Um I don’t know if you if you use Claude. Some people are using Claude here. Uh but in in Claude, you might take your previous conversations and try to refine all your workflows and your skills based on the things that you’ve done in the past. So, this is something that’s makes

27:51 sense to to save in the long run, right? You also have things like procedural memory. Previous workflows that have worked very well for your system. For instance, you worked on this front end, and then you created a very beautiful design that you like. You might take the whole conversation and turn that into a workflow that is repeatable and re- reusable. So, that the next time you’re working on a front end, the results will be similar to the previous one, right? So, these are the things that we we will

28:22 see on the workshop. And some people say, “Okay, why do I even need all of this?” Like people that are very that have animosity towards agent memory. People say, “Okay, let’s just put like 15 million uh context window.” Even though it’s not not possible yet, but some people really believe that this is the the thing, right? But the context window is a type of short-term memory. So, it’s useful for some things, but not for all of them. And one of the problems that that happen

28:53 with with working with a with a context is this thing called context rot or context degradation over time. And what happens is that the more things that you put into the context window, the less attention there will be for each one of the things that are in the context. So, at the beginning of a conversation, and this is a a a famous problem that that the context window has is at the beginning of the conversation, it will stay on track a lot because

29:24 you are just you just started the conversation. So, let’s say that for instance, like in school, right? Or if I’m having a conversation with you, um I might have the a chat with you for 30 minutes, and your attention to me is very very high because I just started speaking. But if the conversation goes on for 8 hours, then you want to punch me, right? Because I haven’t shut up, and you haven’t learned almost anything at the end. And the the problem is that attention, like us humans, is very limited. So, the

29:55 more things that you put in the context window, the attention matrix of the neural network will all also degrade, and it will like scale quadratically because the attention matrix, you know, is one token is essentially a reference of one token for every other token in the context window. So, the bigger the context window is, the matrix scales on the number of rows and on the number of columns as well, which is a problem. So, you want to keep the context window as small as possible to avoid context

30:27 rot. And memory engineering, the components of memory engineering, um so, it’s like designing, building, and doing everything around building agent memory for AI agents. And we want to retain, recall, reuse, and refine this data in some type in some way. So, it is a discipline, right? And here we have Valentin, for instance, and we have people from Oracle,

30:57 my colleagues all over the the room. So, if you see them, you can say hi to them. Valentin here, he’s working on the development or he worked on the development of this agent memory package. So, if you have any questions about this, you can ask him. Um OAMP or Oracle Agent Memory Package is the managed answer that we have in Oracle to lots of the problems that you will find when working with this type of data. For instance,

31:28 as an engineer, if you do not have a managed solution, you need to make a lot of decisions. You need to see, when do I do context compaction? When do I summarize? How do I write it the summarizer? What do I keep? What do I not keep? Um what do I extract from my previous conversations and when? How many tokens do I use for this problem? And lots of these things, right? With OAMP and Oracle Agent Memory

31:58 Package, you can actually just do all of these in one single line of code. And we want to make it easier so that we reduce the cognitive load of AI engineers and AI agents as well. And with this context card thing, you will get something like what you see on the left. And this is kind of an explanation of what each part does on it. But essentially, the the topics that you see, for

32:29 instance, from any conversation, you can create a context card from the thread or from the conversation. The topics will orient the model. The summary will compact the thread and it will state the current intent of an AI agent. And the relevant information has three different parts, which is the the facts, the preferences, and the memories that are associated to this conversation. Uh then you have the episodic memories that explicit explicitly track the

33:01 unanswered question that is going on right now. And the recent messages give like local context to the model. So, whenever you’re feeling like unsure, what do I need to do right now with the data that I have or this conversation, you might use the Oracle agent memory package on on Python and it is all assembled by one single call. Yes. >> So, you’re not suggesting that this goes directly into the model. This is actually something that is used by the furnace

33:32 >> Exactly. >> to harness what’s going into the model. >> Yes. Yes. >> And this is the component that you’re talking about that Oracle uses. It knows about the structure and it uses this structure to >> Exactly. Exactly. So, this structure This is an abstraction that we build on top of the model because the model we have no control over in most cases, right? Um The thing is that we can build on top of that whatever abstractions we want to make the model perform as reliable as reliably as possible. Yep.

34:04 So, the memory and the semantic layer they kind of work together with I know we talked only about memory components. Uh I’m going to walk you uh quickly because I’m I don’t have a lot of time um to the semantic layer, to the semantic components of it, right? And the semantic layer is the the meaning of what’s going on behind it that you kind of assume that that happens, right? So, anyone from Germany? Okay, so I apologize for my pronunciation, but I’m going to try. So,

34:36 the Umwelt is like the ambient of a model, and this was coined by Jakob von Uexküll. Um and this guy said that essentially every organism in the world uh that is living perceives its reality through a lens. And the lens is what it has access to. For us humans, for instance, we have our eyes, our senses, right? So, everything that we perceive and everything that we live, all our experiences, are seen

35:07 through this lens, right? And an agent doesn’t have human like senses. But it has also a kind of semantic layer or a semantic lens that everything that you ask it is filtered through. And this lens is essentially what you train it with and then what you also give context to. So, everything that you talk to the agent, right? When you talk to the agent, this agent will look it through the Umwelt.

35:39 And the semantic layer is essentially the agent’s Umwelt. So, the Unsaid is what what what I want to focus on. Like, for instance, organizational knowledge or enterprise knowledge. Things that when you’re working with a colleague you don’t mention this because this is already you know, known between you and your colleague. You don’t need to specify everything that you work on every day, but if there was someone else, like a a child, that wanted to that [clears throat] wanted to start uh

36:10 working with you you would have to specify everything very very in detail. These are all the Unsaids, and this is what the semantic layer captures. So, the Unsaid is the actual tribal knowledge that that an an enterprise has, right? Institutional knowledge as well, like how the data is modeled, how the queries are executed, what is the metadata. All these things are in the semantic layer. Um and quickly

36:40 just for you to know that we’re also going to implement a very minimalistic agent loop. And an agent loop is like the driver of the model, right? It lets a model be kind of independent and autonomous. And it is what makes a model it it it turns a model into an agent, right? And this is like the simplest agent loop that you can find. It’s kind of this observing and reasoning and then acting

37:11 part that happens all of the time, all of the time. And you know, it’s it has to be failure failure resistant so that we never exit the loop. This is the idea. That the agent you give autonomy to the agent. Well, you give autonomy to the model so that it becomes the agent. And then on the context engineering part, which is the last part of the seven layers of an agent harness, you also have lots of things that you can do. For instance, the toolbox pattern

37:43 and the skillbox pattern, which we are This is on the workshop as well. And these are ways in which you can store the available tools and the available skills of a model so that they are retrieved optimally. And what you want to do is only retrieve the tools and the the skills when they are actually needed and put it on the context window only when needed. Every iteration of an agent loop, you will see if this is actually the right

38:13 place and then if it’s not, you can just take them out temporarily. So, uh Oh, sorry. I thought I I saw a question. So, we will see all of these in the workshop. I don’t want to take up too much time. But, the idea is that we will assemble the context at every iteration on the agent loop. And then, continual learning is the part that that we talked before about the ability to get better over time with the things that we’ve done with the

38:44 model, right? So, a frozen model, as we saw, it doesn’t get better, right? But, there are ways in which we can make an agent improve in the weights, which is the part that we are going to work on. Uh, in the representation part, so on the embedding and the re-ranking part, and also in the context window. And these are the three types of continual learning techniques. We are going to focus on the workshop on the context and

39:14 the token space because it’s one of the easiest ones and all one of the least expensive ones, as well. I assume no one is a millionaire or not many of us are millionaires. So, this is also the most achievable and the most realistic way to change the model behavior over time. So, without further ado, I just want to introduce you to this part, which is skill promotion and workflow workflow

39:45 promotion. So, those skills or those workflows that you’ve done for 3 4 hours, right? You’ve been working for the whole day on a workflow. You were able to successfully do your job, um, and then these things can actually be retrieved. They can be stored into the memory components that we’ll see. And then, we will see about skill promotion. So, if we promote a skill, for instance, we can promote it through a distillation process and

40:16 create a better skill.md than the original. So, we will retire the old version and update from the new version. And this allows us to do some kind of continual learning on our own skills that turn them into more customized skills for ourselves, for our tone, our way to work, our preferred like our preferences like for instance, let’s use this specific library because I really like the look and feel of it. Let’s use this specific

40:48 database engine because it has less bugs or I found it easier to work with. All these things can be promoted into reusable and improvable skills over time. So, this is what the whole harness would look like at the end of the of the workflow. Sorry, at the end of the workshop. And hopefully what you leave this you’ll leave today with a better understanding of all the specific components that make up an agent

41:18 harness. So, let me go to here before I forget. If you are interested in the Oracle agent memory package or are working on the agent memory package, we have a Discord server in which you can just chat with us as staff. If you’re a Discord user, just feel free to to join this Discord server. I’ll put the link later as well. But, without further ado, let’s begin with

41:50 the actual workshop and I will tell you how. So, let me show you first what we are going to build, right? This is an upbook. Oh, sorry. Uh you don’t see this. Yeah. Yes. Uh the the the first Yes, the workshop instructions, right? So, for those of you who weren’t here,

42:20 you can just go into this website, register with your GitHub user, and then you will get an invitation like this to a GitHub repo. And from here, we will create a GitHub code space. Uh so, please make sure to do that right now. And all my colleagues are are around the room room to answer any questions that you might have during the creation of the code space, etc. Yes. Yeah.

42:55 The internet? Are you having issues with the with the internet? Okay, let’s see. So, do you guys have my the same Wi-Fi? Um ai.engineer Wi-Fi? Okay. So, can someone assist people with the Wi-Fi if possible? Yeah. Please fix the system, whoever. Do it. I can’t I can’t.

43:28 Yeah. Yeah. Uh my colleagues will will take a look at that if Right. Yeah. Yeah, it it always happens, you know? Yeah.

44:00 All right. Say again, sir? Yes, they will go into the air I >> believe they will they will go into the AI engineer. So, if you go into the session uh I’ll make sure to to to go there. If not if you either join the Discord or any other, you know, you can just message me as well on LinkedIn. I’ll gladly give you the the slides if you want. All right.

44:30 Um video team Can you turn me on? Uh sorry, switch me on. Please. Uh Oh, perfect. Thank you. So, what I’d like to to show you is that we have built also, apart from the notebook that we’re going to go through we also built this app book. And with the app book, you can actually test every of the individual components of an agent harness

45:02 uh individually, right? So, just for me to show you that this is possible. And you will get this automatically deployed in GitHub code spaces as well. So, you will have um this uh already deployed. And you’ll say, “Well, how am I actually making requests?” We are going to be making requests to uh Oracle uh oh Okay, not fun. Demo time. Um the idea is that

45:32 Yeah, okay. I I know what’s happening. So, I lost connection to my to my code space because of inactivity. Let me restart. Uh this app book is going to allow you to create and chat and interact with the whole agent harness. And the models that we’re going to use are actually deployed on a managed service that we have on Oracle called OCI the generative AI service. We have partnerships with Google, with Meta and with OpenAI and with xAI for the

46:04 time being and we can actually provide inference to their models through our managed uh service. So, just think of us as the enterprise open router, if you’d like. So, let me just go so you get started uh you can get started. This is the repo, right? So, the agent harness workshop. If you’re here, um thank you for starring that, by the way. Um if you’re here, you just have to click on open in GitHub code spaces.

46:36 And it will take you here. And you can select as many cores as you like. Uh if you So, if you want to create this with eight or 16, please don’t because I I don’t I’m paying for this myself. But uh you might also create this with more resources. But just create the code space. And I’m going to pay for it, as I said. So, don’t worry about that. And this will create a new a new code space instance. And once it

47:06 finishes, which it hasn’t yet, I will show you what we can do with the app book and the notebook. But the idea is to use the remainder of the time the we have 1 hour and 15 minutes to go through the notebook. And you will actually have to Let me show you on on GitHub, actually. Um you can go here. And inside the notebook, after you deploy the code space,

47:37 you will get a a student notebook here. And this is one part of the of the workshop, right? And here we’re going to implement the whole agent harness substrate from scratch. So, we’re going to start with only the model, and then we’re going to keep adding layers to the agent harness, as we saw the seven layers, right? And we’re going to be here to to assist you. You will have to do some to-dos. So, let me show you. There are a couple of things to do for

48:08 you. So, for instance, the first thing that you need to do, you need to create a question, right? The simplest thing of everything. You just have to communicate with a model with no agent harness implemented, right? So, the first thing you’ll need is to ask any question that you like. This will go through the OpenAI completions API, and it will return you a response. This is the simplest of all. And then, we will start adding search, retrieval, encoding,

48:41 and all the other components that we that we have seen. There are a total of 19 things that you need to do. If you finish first, raise your hand, and I will give you a hug, because I don’t have anything else. And yep. So, anyone already deployed that the code space? Okay, one person. Okay, good job. So, any Yeah, if you have any questions or any problems, uh let me know. But, this

49:12 is what it looks like when you have it deployed, okay? So, let me go through this quickly. So, you will get an app, right? Um and the app will already have everything that you need. If you want to deploy this app yourself, you might change this total recall port here. Let me show you how I did it again. I go into ports.

49:43 I clicked on the visibility of the port, and I changed this to public. And then, this is now using a public gateway so that if I open the browser, I can actually get access to my individual total recall instance. So, for instance, if I ask a question like show the total revenue by product category. And of course, this is mission control, so this is this has all of the components that we’ve spoken about

50:15 implemented already. You will get also a context window visualization of the things that are going on on the background. For instance, these are the tools that were selected by the agent harness to be loaded into the context to answer this question. This is the schema that’s happening, and then we can also take a look at the individual agent traces that are going on. For instance, which skills which skills are being loaded, what sources of data are we taking and what are the tool calls being used

50:46 like for instance running SQL commands etc. to answer your question. So, the question is still being built. It’s taking 16 steps. And you know, it’s going to for instance here, it detected an error, right? But because our agent harness is uh fault tolerant, it will keep trying because it’s it has an agent loop implemented etc. right? So, all these things will actually yield you this result uh from from the data in the database, right? And you can actually go into the

51:19 context window, see how many tokens were using. And if you’re particularly interested in some of these parts for instance, the Oracle agent memory package for instance, you can interact also with only the the context card, how to how the context card is being created etc. etc. So, you will all get this deployed in your code base. Yeah. >> So, you showed that the harness is choosing most

51:50 >> Yeah. So, his question for those of you who didn’t listen where what happens if you have thousands of tools in an organization, right? Well, we introduce this concept called the toolbox pattern in this course with Andrew Young. And the thing is that you can optimize so that the retrieval of these tools is

52:21 negligible. So, you will use hierarchical navigable small world indexes that use a graph a graph structure and then each node in the graph is a vector index or a vector store. And then you can actually like HNSW indexes, they can be created created for these types of problems only in the data is not not in files. So, great question. It doesn’t have to worry you until you reach

52:53 millions and millions of users. And tool calls like different specific tool calls, you might not get 5 million. It’s more like reading a file, writing a file, grabbing and all these kinds of tool calls that that that we do every day, they typically don’t excee- like exceed 100 or 1,000. Um but by being on a vector store, you abstract the complexity the complexity and the amount of it. You can just make a query to 1,000

53:25 just as simply as you would 10,000 because of the storage component that we we choose it, which is an HNSW index. >> So, let’s assume that maybe two different organizations within the organization, business units. They have access to something let’s say like >> Yeah, some of them they have access to confidential data for instance. Some of them don’t.

54:01 Yes. Great question. So his question was what happens if the tool descriptions that two different companies have are very similar, right? And one of the things we we can do on the toolbox pattern is actually generate with LLM enhanced toolbox descriptions for this for specific tools to increase the separability of the of the tools. So if you think that the current descriptions of a tool or of a skill as well are not

54:33 enough, you can actually enhance them with LLM retrieval like you would uh instead of running for instance named entity recognition, which is very uh caveman style, you can also do something more sophisticated, which is enhancing the the kind of like docstring uh enhanced representations of a tool so that you increase the separability when you’re doing vector search. Does that answer the question?

55:08 The agent loop? Is that Oh, yes. Yes. That’s Yeah.

55:38 In your harness, how do you Yes. So, there there is a there is a limit, of course, because we don’t have infinite money. So, we can’t just keep trying and trying over and over if the generations are just hallucinations, right? There is a cutoff point that I set depending on the frontier LLM that I’m using. For instance, for 4.1 fast reasoning, which is the one that we’re using here,

56:10 I found that a value of 8 to 12, like maximum number of tool calls before giving up, is correct. Depends also on the on the accuracy and the and the correctness of the model. Like, for instance, in this case, it was just able to to to to show it in two. Before, it was able to find it in 16. So, sometimes it will have a faster retrieval. Sometimes, you will need to be a little bit more patient. But, what I like to define is a

56:41 variable like a hysteresis variable that holds the amount of patience that the harness will get with the model. Then, you can do some other things. Like, for instance, if the model is garbage, you can just use or use it like a router for more different like for difficult types of problems, you will route this problem to a frontier LLM. And then, for the easier types of problems, you can just attach an open weights SLM, for instance, which will be more more

57:11 interesting. Yes. Yes. I think that’s Like, my personal opinion is that the future is mixture of small experts for for each type of problem. Some companies that they have developed like 100 million parameter models that work exceptionally well for one type of problem. And if you just have an aggregator and an orchestrator that routes the correct model to that like

57:44 the correct query to that model, then you will have a very token efficient type of agent harness. So you you you can actually do model routing inside the agent harness. Some companies are actually essentially only doing that and they will charge you like let’s do I’m going to charge you 10% of the tokens that I’m going to save you from the original amount of money that you were going to spend. Right? Uh so let’s let me show you the the student

58:15 notebook, right? So once you are inside the student notebook for those of you who are not familiar with Visual Studio Code, you might need to select a kernel here so that you run the the notebook. So you might select Python 3.12 here and then you can just start reading. If you stumble into a to-do that you need to do, you have a docs folder with all the explanations, the individual explanations that you need

58:46 to solve this specific problem. For instance, the first to-do, which is just talking to the reasoning core, to the model layer without doing anything else, will just explain what you need to implement on that cell so that it works and you can proceed to the next one. And also have the solution. But if you if you’re not lazy, you will try. And I hope that you that you try. And we will be here answering questions around the room. I’m going to turn off my

59:19 microphone, just come down with my colleagues, and then let’s chat about it for the remainder of the of the session. And if you have any questions or you like to talk um more to us, please come by and and swing by the booth, the Oracle booth. We’ll be there every day uh all the time. And you know, it makes it makes us feel good like we are wanted and we have friends. So, if you want to come up to us, just chat with us a little bit. It It

59:50 will be nice. Internet, how’s the internet? I feel like Caesar.

1:00:20 All right, so I’m going to leave this here. I’m going to keep this here and I’m going to come down. >> [music]