README.md

status	active
portfolioState	deployed
updated	2026-10-30
section	Agents on GitHub
audience	platform-engineer team-lead developer
level	advanced
duration	55
format	core-talk
decision	What can agents run continuously in my repository, and how do I keep them in bounds?
prerequisites	copilot-primitives
related	copilot-on-github copilot-hooks agentic-sdlc
references	url label verified https://github.github.com/gh-aw/ GitHub Agentic Workflows docs home 2026-10-21 url label verified https://githubnext.github.io/gh-aw-wizard/ GitHub Agentic Workflow Generator (wizard) 2026-10-21 url label verified https://github.github.com/gh-aw/patterns/issue-ops/ IssueOps pattern 2026-10-21 url label verified https://github.github.com/gh-aw/blog/2026-01-12-welcome-to-pelis-agent-factory/ Welcome to Peli's Agent Factory 2026-10-21 url label verified https://github.github.com/gh-aw/introduction/overview/ GitHub Agentic Workflows overview 2026-10-21 url label verified https://github.github.com/gh-aw/introduction/architecture/ GitHub Agentic Workflows security architecture 2026-10-21 url label verified https://github.github.com/gh-aw/reference/frontmatter/ Frontmatter reference 2026-10-21 url label verified https://github.github.com/gh-aw/reference/triggers/ Triggers and repository access roles 2026-10-21 url label verified https://github.github.com/gh-aw/reference/command-triggers/ Command triggers 2026-10-21 url label verified https://github.github.com/gh-aw/reference/engines/ AI engines 2026-10-21 url label verified https://github.github.com/gh-aw/reference/tools/ Tools 2026-10-21 url label verified https://github.github.com/gh-aw/reference/network/ Network permissions 2026-10-21 url label verified https://github.github.com/gh-aw/reference/safe-outputs/ Safe outputs 2026-10-21 url label verified https://github.github.com/gh-aw/reference/staged-mode/ Staged mode 2026-10-21 url label verified https://github.github.com/gh-aw/reference/copilot-cloud-agent/ Safe outputs for the Copilot cloud agent 2026-10-21 url label verified https://github.github.com/gh-aw/reference/integrity/ Integrity filtering 2026-10-21 url label verified https://github.github.com/gh-aw/reference/cost-management/ Cost management 2026-10-21 url label verified https://github.github.com/gh-aw/reference/outcomes/ Outcomes 2026-10-21 url label verified https://github.github.com/gh-aw/experimental/trace-graders/ Graders (experimental) 2026-10-21 url label verified https://github.github.com/gh-aw/experimental/experiments/ A/B experiments and evals (experimental) 2026-10-21 url label verified https://github.github.com/gh-aw/setup/cli/ gh aw CLI commands 2026-10-21 url label verified https://github.github.com/gh-aw/patterns/chat-ops/ ChatOps pattern 2026-10-21 url label verified https://github.github.com/gh-aw/patterns/research-plan-assign-ops/ ResearchPlanAssignOps pattern 2026-10-21 url label verified https://github.github.com/gh-aw/reference/faq/ GitHub Agentic Workflows FAQ 2026-10-21 url label verified https://github.com/github/gh-aw/releases gh-aw releases 2026-10-21 url label verified MSBart2/FanHub#202 FanHub /triage run on issue #202 2026-10-21 url label verified MSBart2/FanHub#193 FanHub issue-to-draft-PR lifecycle pilot 2026-10-21

GitHub Agentic Workflows: Continuous AI for Your Repository

The Question This Talk Answers: "What can agents run continuously in my repository, and how do I keep them in bounds?"

Duration: 55 minutes | Audience: Platform engineers, maintainers, and team leads who own a repository's .github/workflows

Thesis: An agentic workflow is a readable contract: frontmatter grants authority, safe outputs are its only write path, and the source, compiled lock, and run audit expose the authority chain.

Present and Study

• Open the deck: 38 animated slides. Press N to open speaker notes beside it.
• Speaker notes: talk track, questions to prepare for, demo cues, first-party evidence, and a time budget for every slide. The notes print as a handout.

The deck and notes live in slides.js. Edit talk content there; this page stays a short landing page.

1. Why and When

Use the recurring + judgment + bounded-write test to choose work. Issue triage is the running example; a gallery shows event-driven and scheduled patterns, with a human owning each handoff to implementation.

2. Read the Contract

Read the real FanHub workflow source, body, compiled lock, safe outputs, and staged mode. The authority chain stays visible from the declaration through the run audit, including its deliberate trust choices and boundaries.

3. Build One Live

Use the wizard to turn task intent into a prompt, then let repository context inform a draft. The reviewer narrows allowed outputs, limits, target, and staging before compiling and inspecting a preview.

4. Run and Judge

Separate job status from domain results. Use run Summary, gh aw audit, gh aw logs, experimental evals, and Outcomes for distinct evidence, know when another tool fits better, and tell Copilot automations apart from agentic workflows.

What You Can Do Today

Today

• Try: Choose one recurring repository task and write down its judgment, initiator, and smallest useful write.
• Expected signal: A wizard prompt that states the task, trigger, and output without granting unrelated authority.
• Validate: Review the prompt with the five questions in the speaker notes before copying it to a coding agent.

This Week

• Build: Draft the workflow in a test repository, compile the lock, and enable staged mode.
• Expected signal: A staged preview that a maintainer can compare with their own expected labels or comment.
• Validate: Inspect the run Summary and gh aw audit; compare the proposal with a maintainer's decision.

This Month

• Pilot: Run staged triage across representative issues and collect corrections, accepted outputs, and run cost.
• Success measure: Agreement with maintainer decisions and accepted Outcomes; treat any target as unmeasured until the pilot produces evidence.
• Boundary: Pause when evidence is weak, output exceeds the declared boundary, nobody owns review, or the task is deterministic.

Apply It to Your Work

• Candidate task: A recurring issue, failure report, backlog digest, or research-to-plan handoff in your repository.
• Decisive context: Relevant code, issue history, existing labels, repository conventions, and the team's trust and network policies.
• Delegation and authority: Let the agent read and propose; a separate job applies only declared writes; a named maintainer accepts, adjusts, or stops the result.
• Evidence: Inspect the staged output, compiled permissions, audit, and later Outcomes; narrow the contract or choose another tool when evidence fails.

Key Artifacts

• examples/example-flow-prompt.md: saved wizard prompt and its FanHub intent.
• examples/triage-wizard.md: repository-aware workflow example, including frontmatter and judgment instructions.
• examples/triage-wizard.lock.yml: compiled lock used to inspect jobs and permissions.
• examples/fanhub-issue-research.md: supporting repository research.
• examples/fanhub-intake-pilot.md: supporting pilot context.

Related Patterns

• Copilot on GitHub: coding agent, code review, and Code Quality after planned work is handed off.
• Copilot Hooks: deterministic controls around agent sessions.
• Agentic SDLC: repository readiness and the path to more autonomous delivery.

References

First-party sources for each slide claim are linked in the speaker notes. The full reference inventory remains in this page's frontmatter.