How to move AI from code completion to agentic 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: How to move AI from code completion to agentic workflows · Channel: GitHub · 01:39 · Watch on YouTube

Topics: AI agents, Software engineering with AI, AI governance and regulation

Summary

Developers often begin using AI for code completion and questions about their codebase without progressing to broader workflows. The video contrasts adding AI to an existing process with redesigning the process around an agent. In GitHub’s example, an agent checks merged feature pull requests, decides whether documentation is needed, drafts it, and requests review from the engineer who shipped the feature. The recommended approach is to start with a desired outcome, provide relevant context and governance, and expand an agent’s role incrementally as its capabilities become clear.

Key points

AI use often remains limited to code completion and answering questions about a codebase.
Adding AI tools does not necessarily change who directs, performs, or assesses the work.
In the earlier documentation process, writers noticed features after they shipped and tried to reverse engineer the changes.
The redesigned workflow starts when a feature pull request merges.
The agent reads the diff and linked issues to decide whether documentation is needed.
When needed, the agent drafts documentation using the existing style guide and opens a draft pull request.
The agent requests review from the engineer who shipped the feature.
The team sets standards up front and assesses the agent’s output at the end.
Start with the outcome, then expand agent responsibilities gradually based on reliability, context, and governance.

Chapters

Start	Chapter	Summary
00:00	Moving beyond basic AI assistance	Developers often use AI for code completion and codebase questions but may not move on to broader workflows.
00:35	Redesigning documentation with an agent	A GitHub team replaced an after-the-fact documentation process with an agent that responds to merged feature pull requests.
01:08	Expanding agent roles incrementally	The team sets standards and reviews outputs, while the recommendation is to start with a clear outcome and grow agent responsibilities carefully.

Quotes

moving from seeing it as a toy to a tool. — 00:00

Many organizations add AI tools without reconsidering who directs, performs, and assesses the work. — 00:35

The agentic version redefines the workflow itself — 01:08

Start small, build a sense of what agents can reliably do — 01:39

Concepts and entities

code completion, agentic workflow, pull request, governance constraints, GitHub

Full transcript (original language)

00:00 There seems to be a gap. When developers first start working with AI, moving from seeing it as a toy to a tool. [music] They’ll typically start with things like code completion, asking questions about their code base, and then not necessarily moving beyond that. How do we support our developers moving to that next step in a way that goes beyond just yet another enterprise-level mandate? >> Many organizations add AI tools without reconsidering who directs, performs, and assesses the work. Let me highlight an

00:35 example we’ve shared on our GitHub blog. One of our internal developer tooling teams used to run docs the traditional AI on the side way. An engineer ships a feature in the product repo, later a docs writer notices, opens the pull request, and tries to reverse engineer what changed. AI, if it was used at all, was there just to help write or tidy the draft. The team rebuilt that process with the assistance of a GitHub agentic workflow instead. The moment a feature pull request merges, an agent reads the diff and the linked issues, and decides

01:08 whether it needs documenting. If it does, it drafts the doc against their existing [music] style guide, opens a draft pull request in the docs repo, and requests review from the same engineer who shipped the feature. The agentic version redefines the workflow itself, and the team members’ role narrows to setting the standards up front and assessing the output at the end. I recommend starting with the outcome you need, and then carefully examining what an agent can do given the shared knowledge or context you can provide,

01:39 and the governance constraints you can set. Start small, build a sense of what agents can reliably do, and under what conditions, then incrementally expand their role across the rest of your life cycle.