MKLab · tech talk
Agents in your repository
GitHub Agentic Workflows: automation that reads, thinks and writes back. Triggered by events, not chat.
Space to begin
I · The question
Everyone is asking for one.
I need an agent, something that automates tasks for my team with little or no interaction. The question comes up everywhere. And if your team lives on GitHub, already running GitHub Actions for CI, the answer is standing closer than you think.
II · The map
Three places an agent can live.
An assistant in your chat and office apps answers you. An agent beside you in the IDE or terminal works while you watch. An agentic workflow lives in the repository and works when events happen, whether you are there or not. One word of discipline: an agentic workflow can act as an agent, but agent is the bigger word. This talk keeps the two apart.
III · Nobody chats with it
So how does it know?
There is no chat window. The instructions are written down in advance, and the trigger replaces the conversation: an issue opened, a pull request, a comment, a label, a schedule that reads like prose: weekly on monday. The event brings the context. The file brings the goal.
IV · The boundary
It only hears GitHub.
An agentic workflow wakes on GitHub events, and on nothing else. A new issue in an on-prem Jira, a push to a local GitLab or Bitbucket: silence. If your work lives elsewhere, bring it to GitHub or bridge it with webhooks, each with trade-offs. Fully on GitHub, this is brilliant. Half elsewhere: know the gap, and plan the bridge.
V · Beyond CI
Same runners, new kind of work.
A classic GitHub Actions workflow runs the same steps every time, which is exactly what you want for building and deploying. An agentic workflow gets a goal and judgment: it reads the issue, investigates, decides, and writes something back. Your pipelines are not being replaced. They are getting a colleague.
VI · The file
Written in markdown, run by Actions.
A file in the repository: frontmatter for the trigger, the permissions and the engine, then plain prose telling the agent what to accomplish. Compiled into a locked Actions workflow. The engine is your choice: Copilot, Claude, Codex, Gemini or Pi, and the file reads the same whichever one thinks inside it.
VII · Guardrails
Powerful, and fenced.
The agent runs read-only in a sandbox behind a network firewall. It cannot push to your code. It proposes: an issue, a comment, a pull request, through safe outputs that separate jobs validate and a human reviews. The guardrails are not a limitation. They are what makes the rest usable.
VIII · Where it helps
Small chores, done nightly.
The wins are not moonshots. Triage the new issues. Investigate the failing build at dawn. Summarize the week. Tend the backlog. Keep the docs honest. Recurring work with judgment in it, small enough to trust an agent with, valuable enough that it compounds.
IX · The whole team
Not only for developers.
The developer gets a nightly investigator that turns broken builds into annotated issues. The tester gets a reviewer that reads merged changes and proposes missing tests. The project manager gets a Monday morning status woven from the week's work. The product owner gets feedback labeled, deduplicated and summarized into the backlog.
X · Early days
Early, and worth being early to.
GitHub marks Agentic Workflows as a public preview: it moves fast, details change, runs cost Actions minutes and AI credits, and the output deserves human eyes. That is not a reason to wait. It is the best moment to learn: start small, read what it writes, and let trust and responsibility grow together.
XI · The sooner the better
This is where automation is heading.
Agents that live with the work and act on its events, this solution or ones like it, are the future of team automation. The sooner your team runs its first workflow, the sooner the benefits compound. Pick one recurring chore, one markdown file, and begin.
MKLab · EST. MMXXV
That was agentic workflows.
This is the end of the presentation. Thank you for watching. The talks on MCP and on teaching your Copilot pair well with this one.
