Blog

Agent of the Day – November 16, 2026

Some workflows exist to watch other workflows. Today’s spotlight is one of them. It checks the health of agent jobs on a daily schedule and reports what it finds in a GitHub Discussion. Its recent runs all finished cleanly, which makes them a good place to look at a quieter question: how much work does a steady monitor really do?

Agent of the Day: Agent Job Health Monitor

Section titled “Agent of the Day: Agent Job Health Monitor”

Agent Job Health Monitor runs on a daily schedule. Each run produces one discussion, and every run in this window succeeded and created exactly one.

This is the first time this workflow has been featured. The rotation moves through active workflows from the least recently featured onward. The workflow ahead of it, ace-editor, had no runs in the last three days, so it could not be judged on live evidence and the rotation moved on to this one.

RunDate (UTC)DurationTurnsAudit tokensDiscussion
43525284919Nov 1321.6m603.25M#77583
43651937051Nov 1426.5m3547k#77876
43773614833Nov 1524.1m2339k#78176

Token figures come from the audit output. The log summary counts tokens differently, so compare the audit numbers with each other rather than against other tools.

The November 13 run is the outlier. It took 60 turns and about 3.25 million tokens, roughly ten times the November 15 run’s total. The audit classed it as a resource-heavy run for an issue-response task and recommended trimming turns, tools, or write actions. The later runs took two or three turns, yet they still ran for 24 to 27 minutes, so turn count does not explain duration on its own.

Every audit in the window carries the same two assessments: “resource heavy for domain” and “poor agentic control.” These are judgments, not failures. All three runs completed, reported no errors, and made two write actions each.

The November 15 audit includes several checks that deserve attention:

  • Loops grader: the loops check failed, reporting 275 against a threshold of 3. The run recorded zero retries, so this is a signal to investigate, not proof of a retry storm.
  • Context growth: the context grew about 29 times over the run. That growth is one reason the token total is high for a run of only two turns.
  • Cache window: the audit warns that gaps between turns may exceed the five-minute prompt-cache lifetime. If they do, the context is re-read at full cost.
  • Network activity: the firewall logged 155 requests and blocked none.

The chart below is the repository’s general activity view, shown for context. It is not a chart of this workflow’s runs.

Agentic workflows activity

A monitor that succeeds every day can still be expensive. The clearest lever in this data is the turn count. The November 13 run shows how a long run multiplies cost without changing its outcome. The open question is whether that run was an outlier or an early sign of a pattern that the next few weeks of runs will show.

Read the workflow source to see how it checks agent jobs and reports results. To build something similar, start at github/gh-aw.

Weekly Update – November 16, 2026

Two releases shipped in github/gh-aw this week, and a steady stream of pull requests landed alongside them. The theme is cost and control: model routing now favors cheaper models for eligible workflows, and several breaking changes need a look before you recompile.

v1.2.3 (November 13, marked pre-release) centers on model routing, safer connectors and merges, and a Go 1.27.4 baseline. Review these breaking changes before you upgrade:

  • Default model routing changed. Eligible workflows that set neither a model nor routing now use goal: cost, mode: auto on GitHub Copilot. To keep the previous selection, set engine.model-routing: false. See the model routing reference.
  • Declarative connectors replace native Jira, Linear, and Azure DevOps integrations. Import .github/workflows/connectors/<provider>.md, choose read profiles and write effects, then recompile. See the connectors reference.
  • merge-pull-request blocks PRs that change protected files. Merge those manually, add protected-files.exclude entries, or set protected-files: allowed.
  • pull_request_target admits same-repository PRs only by default. List exact owner/repo entries under forks to admit forks.
  • AWF and Copilot CLI downloads need a checksum pin. Unpinned and latest downloads now fail closed.

New in this release, you can set model: router, router/economy, router/balanced, or router/robust directly, and create-security-advisory creates draft advisories that are never published automatically.

v1.2.2 (November 10) focuses on work-queue handling, firewall egress, and session normalization.

  • Cancel obsolete work. cancel-work --obsolete adds a read-only dry-run preview (#76180), and queue maintenance now cancels obsolete worker backlog (#76126).
  • Cheaper defaults. PR Sous Chef and other workflows now use economy model routing instead of pinned models (#76325, #75471).
  • Fewer wasted runs. Skip-if-decision gates now cover AWF engine invocations (#76008) and ten more Copilot workflows (#76099).
  • Name-based artifact upserts in update-issue (#75988).
  • Updated dependencies. MCP Gateway v0.5.6 and firewall v0.29.11 (#76311). Terminal safe-output signals are now required (#76194).
  • #78137: Exponential backoff for threat-detect downloads. Release-host 504s used to exhaust six fixed-delay retries in about 50 seconds, which stopped the detector from running at all.
  • #78015: Zip option injection fixed. A staged filename starting with a dash, such as -T, could be read as a zip option. Those names are now rejected.
  • #77989: Firewall v0.29.14 and MCP gateway v0.5.10 become the default pins.
  • #77561: Opt-in trigger reaction removal. Set on.reaction.remove-on-complete to clear trigger reactions when a run finishes. It defaults to false.
  • #77866: Workflow source in gh aw list --json. Tooling can now tell which workflows are managed from an upstream bundle.

Sub-Issue Closer is the scheduled workflow that closes a parent issue once all of its sub-issues are complete. It runs daily on the pi engine with copilot/gpt-5.

Over its last three scheduled runs (November 13, November 14, and November 15), it closed 30 parent issues and commented on each one. All of them were smoke-test tracking issues, such as “Smoke Copilot ARM64 - Issue Group.” Each run took 7 to 8.5 minutes, and all three succeeded.

Its closing comments are a handy audit trail, so if you rely on it, skim them now and then to confirm each parent really was finished.

Usage tip: Use a closer like this for work you track as sub-issues, and review the comments it leaves so you can trust what it closed.

→ View the workflow on GitHub

Review the breaking changes in v1.2.3, recompile your workflows, and tell us what you think. Feedback and contributions are welcome in github/gh-aw.

Ecosystem Spotlight: githubnext/evergreen

githubnext/evergreen is a GitHub Agentic Workflow that keeps selected pull requests green and mergeable. It fixes CI failures and keeps PR branches current with their base. It is aimed at teams that want an agent to repair pull requests they have explicitly opted in, without giving that agent the power to merge or to write to the default branch.

The trigger is deliberately not pull_request. The workflow runs on a 15-minute schedule and on workflow_dispatch. The comment at the top of workflows/evergreen.md explains that generated pre-activation runs can trip confused-deputy guards for bot-authored or indirect PR events. Teams that need immediate handling are told to use a separate deterministic controller that never checks out or runs PR code.

A deterministic gate runs before any agent. A preflight job confirms that the PR is open, carries the opt-in evergreen label and the evergreen_active lease, and still has the head SHA that was selected. A pre-agent step then fetches refs/pull/<n>/head and refuses to continue if the fetched SHA differs from the expected one. A stale checkout fails closed.

A lease prevents two runs on one PR. The workflow creates the evergreen_active label, and it refuses to dispatch the agent if the label cannot be added. A separate cleanup job removes the lease. Runs are queued with cancel-in-progress: false, so a second run waits rather than cancelling the first.

Write rules are operating rules. safe-output-policy.md requires the agent to reload GitHub state after every safe output and to report the operation as blocked if the expected change is not visible. It also forbids direct merges, base-branch writes and branch updates that merge or rebase the base. The workflow enforces the last rule too: a git wrapper placed on the agent’s path exits with status 64 for merge and rebase.

Quota and stop states are explicit. quota-policy.md counts quota per PR, continuing across runs for as long as the label stays. When it runs out, the agent stops, asks for the opt-in label to be removed, adds an exhausted label and leaves one terse comment. Every run ends in one of five named stop states, such as awaiting-controller-recheck or quota-exhausted.

Fail-closed defaults. The workflow ships REQUIRED_CHECKS_JSON as '[]', and a comment labels it a fail-closed placeholder that installers must replace by naming their gates or choosing all-observed mode.

  1. The cron wakes the preflight job, which selects an eligible PR and evaluates the gates.
  2. For an eligible PR, the job adds the lease label and dispatches the agent. The agent job checks out the verified head and runs the Copilot engine with network: defaults.
  3. The agent runs evergreen-diagnose before it edits anything, then evergreen-repair if diagnosis says repair is needed. Both skills are pinned to a commit SHA and installed at runtime, so the prompt does not carry their full text.
  4. Writes are safe outputs: comments, labels, PR-branch pushes limited to PRs with the evergreen label, and at most one review. Each write is verified after the fact, and the run finishes by removing the lease.

Evergreen is installed by an agent, following INSTALL.md. The installer reads the target repository’s guidance and its last 50 closed PRs, proposes a mergeability policy for you to confirm, copies workflows/evergreen.md and workflows/shared/evergreen/ into .github/workflows/, runs gh aw compile and opens an installation PR. After that, apply the evergreen label to any PR you want it to work on.

Before you adopt it, check a few things:

  • The preflight job has contents, issues and pull-requests: write permissions, so review its gate logic whenever you change it.
  • The required-checks list ships empty, so you must name your gates before the agent can clear anything.
  • Public or fork PRs should require trusted approval before any code runs, as the README says.
  • It is early-stage: 13 stars, and its latest commit is from late July 2026.

Even if you never install it, Evergreen is a useful reference. Its gate, lease and verification patterns apply to any agent that writes to pull requests.

Explore githubnext/evergreen, and learn more about building your own workflows at github/gh-aw.

Ecosystem Spotlight: svennijhuis/agentic-workflows

Most agentic workflow examples stop at one trigger and one comment. svennijhuis/agentic-workflows goes further. It is a catalog of eleven installable workflows that take an issue to a reviewed pull request, with a reviewer that can send work back a limited number of times before handing it to a human. It is aimed at teams that want a working, opinionated build loop they can install with gh aw add and adapt, rather than designing one from scratch.

A hard cap on the loop. Each workflow enforces its own limit in its if: condition, so the runner decides whether a run starts, not prompt text. issue-builder.md starts only for a new build or fix rounds 1 and 2. pr-reviewer.md runs review rounds 1 to 3 only, so the builder gets at most two send-backs. The reviewer also derives the round from its own history of Reviewer: changes needed reviews, so a wrong input cannot reset the count.

Review that never merges. The reviewer allows only COMMENT reviews. The workflow explains why: without a separate builder identity, builder PRs and reviews share github-actions[bot], and GitHub rejects REQUEST_CHANGES on your own PR. A maintainer stays the merge gate. The reviewer can add ready-to-merge or needs-help labels, but it never approves, merges or pushes.

Four axes plus a conditional security pass. The reviewer starts standards, spec, deslop and blast-radius sub-agents, each in a fresh context and each asked for path:line evidence. A security sub-agent runs only when the diff touches a trust boundary, such as endpoints, auth, permissions, .github/workflows/*, dependency manifests, secrets, shell calls, outbound network or deserialization. Each axis keeps its own heading and is not reranked, so a reader can see which axis raised a concern.

Fail-closed defaults in shared fragments. shared/security.md closes the network except for gh-aw defaults, requires importers to name any extra host, and keeps threat detection on. shared/defaults.md pins the engine version. max-ai-credits and max-daily-ai-credits use the main workflow’s values when set, then fall back to the first usable imported value. Other limits such as max-turns and timeout-minutes stay in each workflow’s own frontmatter because the compiler ignores them in imports.

Vendored skills with provenance. The shared/skills/ folder holds local copies of eleven skills, mostly from mattpocock/skills. The shared README records the source path, commit SHA and license for each one, along with the update procedure.

  1. aw.yml is the package manifest. It lists the eleven workflows under includes: and sets min-version: v0.93.4, so gh aw add knows what to install.
  2. An issue-builder run implements an issue and opens a builder/* pull request labeled builder.
  3. The builder then dispatches pr-reviewer for round 1. The reviewer posts a COMMENT review that starts with Reviewer: ready to merge. when clean. Otherwise it posts inline comments and dispatches the builder for a fix round.
  4. If round 3 still has findings, the reviewer adds needs-help and posts one summary of open points for a maintainer. It starts nothing further.

From the README, in a target repository:

Terminal window
gh aw add svennijhuis/agentic-workflows/workflows/issue-builder.md@<tag>
gh aw compile

Pin <tag> to a released version rather than a branch. Each workflow names the other in dispatch-workflow, so installing either one brings the other along. The builder also references a CI_TRIGGER_TOKEN secret for an extra empty commit that starts CI. It is a placeholder name, so rename it to match your setup. The install guide covers tokens, secrets and labels for an organization.

Patterns worth borrowing even without the full loop:

  • Put the loop cap in the trigger if: so the runner enforces it.
  • Make the reviewer comment-only when the author and reviewer share an identity.
  • Gate expensive security review on a path-based trust-boundary check.

Explore svennijhuis/agentic-workflows, starting with its workflows README and the builder and reviewer pair. To build your own, start with GitHub Agentic Workflows.

Agent of the Day – November 13, 2026

Some workflows report on the code they touch. Today’s spotlight goes to one that reads its own codebase for patterns worth linting, then tries to turn what it finds into a fix. It is also the workflow that failed on its most recent scheduled run, which makes it a useful case study in how these systems break.

Agent of the Day: Actions and AW Linter Factory

Section titled “Agent of the Day: Actions and AW Linter Factory”

Actions and AW Linter Factory runs once a day. Each run picks one lint rule for the AST checks under pkg/workflow/lint/, looks for real evidence that the pattern matters, and proposes a rule with a fix. It skips a new run while a pull request with [actions-aw-linter] in its title is still open, so it does not stack up competing proposals.

This is the first time this workflow has been featured, so the rotation put it ahead of workflows that had been spotlighted before.

The last three scheduled runs tell three different stories.

RunDate (UTC)ConclusionDurationAudit signal
43267067168Nov 12failure3.4m“Agent reported that it could not complete the task”; threat detection failed
43066035425Nov 11success13.0mHigh token usage (about 995k tokens)
42862531335Nov 10success18.6mHigh token usage (about 1.36M tokens)

The November 12 failure is the one to read closely. The audit classes it as a critical workflow failure, with three errors. The agent stopped before it could finish. The threat detection job also failed, with the reason agent_failure. The audit calls that a tooling error, not a security finding. The failed steps were the “Install AWF binary” steps in both the agent job and the detection job. The run lasted only 3.4 minutes, so this was an early stop, not a timeout in the middle of analysis.

The two successful runs completed in one turn with no reported errors. Their audits still flag two things: high token usage and a “Resource Heavy For Domain” finding. For a daily job that reads a whole codebase, that cost is expected, but it is worth tracking as the rule set grows.

The workflow has already produced a steady stream of pull requests. Each one is titled with the [actions-aw-linter] prefix and is now closed:

  • #69972: Add the daily Actions and AW AST linter factory itself.
  • #71189: Fix workflow_run positive branch cancellation detection.
  • #71929: Add a step-reference lint rule.
  • #73253: Fix release cache lint for Python and Java setup actions.
  • #73832: Fix artifact execution detection for option-prefixed scripts.

The fixes are the clearest sign that the loop works. A lint rule that produces a false positive gets a follow-up pull request that narrows it, rather than a quiet ignore.

Combined workflow activity chart

A linter that writes its own rules is only as trustworthy as its validation. This workflow has a gate that stops it when a run cannot complete, which is what the November 12 failure shows. A stopped run is far better than a half-finished proposal that looks finished. The question for the next few weeks is whether the install step that failed on November 12 is a one-off or a pattern.

Read the workflow source to see how it selects rules and gates its runs. To build something similar, start at github/gh-aw.