Agentic Workflows

September 21, 2026 · View on GitHub

AgentRC uses GitHub Agentic Workflows to maintain this repository and demonstrate an auditable agent loop.

The central demo is:

AgentRC measures the repository
        ↓
gh-aw gives an agent the evidence and a bounded task
        ↓
the agent proposes one instruction improvement
        ↓
AgentRC measures again
        ↓
a human reviews the draft pull request

Current workflow suite

Workflow	Trigger	Maximum output	Purpose
AgentRC Readiness Report	Weekly Friday or manual	One expiring issue	Publish the highest-impact readiness gaps from deterministic AgentRC output
AgentRC Readiness Upgrade	Manual	One draft PR limited to instruction files	Demonstrate measure → improve → verify → review
Issue Triage	Issue opened or reopened	Two allowlisted labels, one issue type, and one comment	Route incoming issues without inventing missing details
CI Failure Doctor	Failed `CI` workflow on `main`	One issue or one comment	Find the first meaningful error and recommend a bounded repair

The imported workflows retain their source: reference to githubnext/agentics. AgentRC-specific changes live in the local source files and should be reviewed when updating from the catalog.

Why these workflows

The selected catalog workflows cover three recurring maintainer jobs:

Issue Triage reduces response latency without editing code.
CI Failure Doctor turns a failed run into an evidence-backed diagnosis.

Two custom workflows make the demo specific to AgentRC:

Readiness Report runs the product deterministically before the model starts.
Readiness Upgrade limits edits to agent instruction files and requires before/after evidence.

The previous daily documentation, code-simplification, and weekly status workflows were removed. They were generic, failed repeatedly in observed runs, or produced recommendations about already retired automation. The smaller suite keeps the ongoing work tied to concrete repository events and AgentRC's own readiness loop.

Safety model

The workflow source declares the boundaries. The generated lock file implements them in GitHub Actions.

All AgentRC workflows use these rules:

Copilot inference uses copilot-requests: write; the agent does not receive repository write permission.
GitHub writes happen through typed safe outputs.
Code changes are draft pull requests.
Scheduled workflows have a stop date, concurrency control, turn limit, per-run AI-credit cap, daily AI-credit cap, and bounded output.
Temporary evidence lives under /tmp/gh-aw/agent/ so it is available in run artifacts.
Workflow definitions, package manifests, generated files, and eval criteria are outside the readiness-upgrade write scope.
No workflow auto-merges a pull request.

See .github/aw/instructions.md for the repository-specific authoring rules.

Five-minute live demo

Use AgentRC Readiness Report for the safest live path.

Before the session

Open the source workflow.
Open the generated lock workflow.
Keep one successful Actions run and its generated issue open in separate browser tabs.
Run the workflow once on the day of the demo.
Record a 60-second fallback showing the run graph and issue.

On stage

Frame the problem: repository instructions drift as the code changes.
Show the source: point to the schedule, read-only permissions, budgets, tool list, and one allowed issue output.
Show deterministic evidence: AgentRC runs before the model and writes readiness.json.
Trigger the workflow: use Run workflow or:
```
gh aw run agentrc-readiness-report
```
Switch to the prepared run: show activation, agent, detection, and safe output as separate jobs.
Open the issue: connect every recommendation to AgentRC evidence.
Show the audit record:
```
gh aw audit RUN_ID
```
Close: AgentRC measures; the agent interprets; GitHub Actions enforces the boundary; a human decides what to change.

Ten-minute technical extension

Run AgentRC Readiness Upgrade after the report:

Show the exclusive file allowlist.
Show why AGENTS.md is explicitly exempted from protected-file blocking.
Trigger the workflow.
Use a prepared run while the live run executes.
Review the draft PR's evidence, before/after readiness report, and validation.
Emphasize that the workflow cannot change product code, dependencies, workflows, or eval criteria.

Local maintenance

Use the stable compiler version pinned in CI.

gh extension install github/gh-aw
gh extension upgrade github/gh-aw
gh aw --version
npm run workflows:compile
git diff -- .gitattributes .github/aw .github/workflows

Commit workflow source and generated files together:

.github/workflows/*.md
.github/workflows/*.lock.yml
.github/workflows/agentics-maintenance.yml
.github/aw/actions-lock.json
.gitattributes

Do not edit generated files or their action pins manually. Dependabot ignores github/gh-aw and github/gh-aw-actions pins because only the compiler should update them.

CI runs npm run workflows:check, recompiles with the pinned stable version, and fails when generated output differs.

Updating catalog workflows

Review upstream changes before updating:

gh aw update ci-doctor
gh aw update issue-triage
gh aw update weekly-repo-status
npm run workflows:compile

Reapply AgentRC-specific budgets, labels, schedules, and prompt edits when an upstream update conflicts. The local source: field records the imported revision.

Demo failure modes

Symptom	Fix
Workflow is missing from Actions	Compile and commit the `.lock.yml` on the default branch
Runtime cannot find a `gh-aw` script	Compiler and generated action pins drifted; recompile the full suite
Copilot authentication fails	Confirm organization billing policy and `copilot-requests: write`
Workflow creates no output	Inspect whether it correctly called `noop`; verify deterministic evidence exists
Generated PR does not trigger CI	Normal `GITHUB_TOKEN` writes do not recursively trigger Actions
Source compiles with shell warnings	Replace quoted command patterns with safe prefixes
Live run is slow	Continue with the prepared run and compare the eventual result afterward

Evaluating value

Do not count runs as success. Track:

accepted versus dismissed readiness recommendations;
draft PR merge rate;
false-positive triage rate;
duplicate failure reports;
reviewer wait time;
AIC per accepted issue or merged PR;
workflow failures caused by stale generated files;
readiness or eval movement after accepted changes.