Working With Kiro workflows

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: Working With Kiro workflows · Channel: Kiro · 06:51 · Watch on YouTube

Topics: AI agents, Task and workflow automation, Software engineering with AI, AI developer tools

Summary

The video demonstrates how to create and run a reusable Kiro workflow for upgrading project dependencies. The workflow analyzes an upgrade, makes code changes, runs tests, a build or type check, and a breaking-change review in parallel, then retries fixes until the checks pass or the attempt limit is reached. The presenter explains how to generate, review, customize, and run workflows in Kiro, and how individual steps can be inspected or steered while the workflow runs. Unlike general agent instructions, workflows explicitly define execution order, parallelism, retry behavior, and the model and thinking effort for each step.

Key points

Kiro workflows turn a repeatable process into an explicitly defined sequence of steps.
The example workflow analyzes dependency upgrades, applies changes, and prepares a pull request.
Tests, build or type checks, and a breaking-change review run in parallel.
If checks fail, the workflow sends observations back to the coding agent and repeats the checks, up to six attempts in this example.
Each step can use its own model and thinking effort, with more reasoning assigned to harder planning and review tasks.
Workflows can be created at workspace level or global level; Kiro also includes built-in global recipes.
The workflow definition can be reviewed in a readable overview or edited as JSON.
Workflows run in the background, and each step has its own agent session that can be inspected or interacted with.
Workflow files can be version-controlled, reviewed, and shared with a team.

Chapters

Start	Chapter	Summary
00:00	Why use workflows	The video introduces workflows as a way to define exactly how Kiro executes a repeatable dependency-upgrade process.
01:05	Generating and reviewing a workflow	The presenter generates a workflow in Kiro IDE, explains workspace and global recipes, and shows how to inspect and edit its definition.
03:14	Running the workflow and its steps	The workflow runs in the background while the presenter explains how to inspect and interact with individual agent sessions.
04:15	Parallel checks, retries, and pull request preparation	The video walks through upgrade planning, parallel validation, retry behavior, the six-attempt limit, and preparation of a pull request.
05:48	Workflows versus agent instructions	The presenter contrasts workflows with general instructions and recaps how to create, customize, run, and share a workflow.

Quotes

That’s where workflows comes in. — 00:00

Starting a workflow doesn’t block the conversation — 03:14

that limit is set to six attempts. — 05:18

those execution rules are explicitly defined. — 06:19

All you need to do is to ask Kiro to generate the workflow, then you review and customize it, then you run it whenever you need it. — 06:51

Concepts and entities

Kiro, Kiro IDE, workflow, Claude 3.5, CLI, Kiro web, recipes, skills, AGENTS.md

Full transcript (original language)

00:00 Intelligent coding agent today can handle pretty complex work. You can give Kiro a prompt or put your team’s process into a scale or agent empty file, and the agent can follow those instructions. But sometimes, you don’t just want to describe what the process should look like. You want to define exactly how it runs. That’s where workflows comes in. Let’s build a reusable workflow for dependency upgrades. Let’s say we need to upgrade dependencies across different project and want to follow the same process each time. And here is the

00:31 workflow we’re going to build. First, we want Kiro to analyze the project that we need to upgrade and figure out what needs to change. Then, it makes the code changes. After that, three verification checks run in parallel. The first one runs the tests. The second one runs the build. And the third one reviews the changes for any breaking issues. If any of these checks fail, it goes back, fixes the issue, and runs all these checks again. Once everything passes, it prepares a pull request. Instead of describing all this process

01:05 again and again across different projects, we will ask Kiro to turn this into a reusable workflow. Let’s get started. We will be using Kiro IDE in this video, but workflows are also supported in other interfaces, such as the CLI and Kiro web. To generate a workflow, I’m using a stronger model with extra high thinking effort. Kiro let you choose from models across different providers, so you can match the model intelligence to the task. Here, I want more reasoning up front, so

01:36 Kiro can decide how to structure the workflow, what should run in parallel, when it should retry, and what model and thinking effort each step should use. I was sending over this prompt, which describe at a very high level the workflow I walked you through earlier. And pay attention to the last part. Um letting Kiro choose the right model and right amount of reasoning for each step. My expectation is that with an intelligent model and right reasoning effort, Kiro will realize that simple checks can stay lightweight, while

02:06 harder planning or reviews can use a stronger model and a more thinking effort. Kiro then generate the workflow while taking into account all the steering files and agent rules you might already have in place. Once the workflow is created, it will show up in the workflows panel under recipes. In this case, I created my workflow at the workspace level, so it stays with this project. You can also create workflows at global level when you want to reuse them across project. You will also see Kiro comes with a list of built-in recipes under global, which

02:40 get you ready to use workflows for common patterns. From here, you can open your workflows and uh review it in a more readable format, including the overview, the individual steps, what comes first, what comes next, what runs in parallel, and where the workflow loops back, things like that. If you want to change anything, you can click edit. This opens up the underlying workflow definition, where you can see the actual JSON, including the required inputs, each step, which agent runs them, which model and thinking effort it uses, and the prompt for that step. Now that we’ve reviewed the workflow, let’s

03:14 simply ask Kiro to run the workflow and fill in the required inputs based on your project. Now the workflow is running. Notice that our main Kiro session is still available. Starting a workflow doesn’t block the conversation, so you can keep asking questions or work with Kiro while the workflow runs in the background. And each workflow step is its own agent session. If I open one of the steps, I can see exactly what that agent was doing and talk to it directly. I can steer it while it’s working or even come

03:44 back after the step is finished and ask follow-up questions about what it did. So, workflows give you a structured orchestration without taking away the ability to interact with individual agents. >> As the workflow is running, let’s take a closer look at what each step is doing. The first thing Kiro does is to analyze the upgrade. This step uses Claude 3.5 with extra high thinking effort because there is quite a lot of reasoning involved. Kiro needs to understand the version change, research the migration

04:15 guidance, find breaking changes, and connect all those back to the code base, and figure out how this repo should be tested and built. It then saves all that planning and findings into an upgrade plan for the rest of the workflow to consume. Next, the coding agent takes that plan and enter the upgrade loop. On the first round, it upgrades the dependencies and apply all the migrations. Then the workflow fans out into three checks that runs in parallel. One runs the tests, one runs the build or type check, and the third one does a

04:47 deeper semantic review for breaking APIs, incompatible dependencies, or migration issues we might have missed. The first two are quite straightforward, so they use lower thinking effort, while the breaking change review needs more judgment. So, that’s why it gets a stronger model and more thinking time. When all three checks finish, the workflow brings the result back together. A lightweight step reads the reports and produce a final verdict. All three checks have to pass before the workflow can move forward. And if

05:18 something fails, the workflow loops back to the coding agent with the observations, and the coding agent will fix them, then the same checks run again until everything passes or the iteration limit is reached. In this workflow, that limit is set to six attempts. And luckily for this workflow, we reached that green check mark on our first attempt. Finally, once everything passes, the workflow exit the loop. The final step pushes the upgrade branch and prepares a a request with the migration summaries, validation results, and

05:48 everything that might still deserve our attention. Now that we have run the workflow Kiro generated for us, you might be wondering, especially if you are familiar with concepts such as skills or agents MD, why not just describe this process there? And the answer is, yes, you can. You can absolutely describe the process to the agent and ask it to follow. But the difference is that those instructions still leave some of execution up to the agent. The agents do decide how to order the work, what runs in parallel, and how

06:19 retries should happen. You will also typically run the task with one model and one thinking effort setting for the whole interaction. But with a workflow, those execution rules are explicitly defined. You can specify what runs in parallel, when to loop back, what has to pass before moving on, and which model and thinking effort each step should use. This level of control ensures a repeatable process every time you run this workflow across different projects. And similar to skills and agent MD, a workflow is just a file. You can keep it

06:51 in version control, review changes to it, and share the same workflow with the rest of your team. To recap, working with Kiro workflows is super super super easy. All you need to do is to ask Kiro to generate the workflow, then you review and customize it, then you run it whenever you need it. Now you have seen what a workflow is, how it works, and how to build, review, and run one in Kiro. I will challenge you to try and build one workflow around a repeatable process you already have today.