Text of github.github.com › gh-aw › blog as Styx crawled it on Nov 18, 2026. The page may have changed since. Only its words are kept, without pictures or layout, and only the first part of a long page.

Open the live page

Blog | GitHub Agentic Workflows

Skip to main content Blog Every day, a workflow in this repository generates software personas and checks how a custom agent responds to common automation tasks. It writes its findings into a GitHub issue. Over three recent runs, the work it did varied a lot, and so did the cost. Today’s spotlight is Agent Persona Explorer. Agent of the Day: Agent Persona Explorer Section titled “Agent of the Day: Agent Persona Explorer” Agent Persona Explorer runs once a day.

Each run creates one issue titled “Agent Persona Exploration” with a Persona Overview and Key Findings section. Its frontmatter also runs an A/B experiment, sub_agent_strategy_prefix_v1 , which compares a per-scenario sub-agent strategy with a batched one. This is the first time it has been featured. The rotation moves through active workflows from the least recently featured onward. The workflow ahead of it, ace-editor , was skipped.

Its two most recent recorded runs, from July 31 and July 13, failed before executing, with zero turns and no outputs, so there was no live evidence to judge. Three runs, three different shapes Section titled “Three runs, three different shapes” Run Date (UTC) Duration Turns Audit tokens Experiment variant Issue 43610365440 Nov 14 14.6m 23 1.44M per_scenario #77806 43730042500 Nov 15 13.9m 20 1.11M batch #78082 43898094066 Nov 16 12.7m 9 690k per_scenario #78410 All three runs succeeded and each created one issue.

The log summary totals 37.1 AI credits across them. Its token counter disagrees with the audit’s: the log reports zero tokens for the November 14 run, while the audit records its highest count. Compare the audit figures with each other, not with the log summary. Token use tracks turns closely. Each run spent roughly 56,000 to 77,000 tokens per turn, so the number of turns is the main lever on cost. The November 14 run took 23 turns and the November 16 run took 9, a difference of more than twofold in tokens.

What the audits flag Section titled “What the audits flag” All three audits carry the same two findings: “High Token Usage” and “Declared Sub-agent Model Not Observed.” The sub-agent model the workflow declares never appeared in the observed run data, so the audit cannot confirm which model did the sub-agent work. Each run is also tagged “Resource Heavy For Domain.” The firewall blocked no requests in any of the three runs.

Each audit also carries a “Partially Reducible To Deterministic” finding. It estimates that 91% of the November 14 turns, 90% of the November 15 turns, and 77% of the November 16 turns were data gathering that a script could do instead. That is the most concrete lead in this data: most turns go to retrieval, not judgment. The chart below shows the repository’s general agentic activity. It is not a chart of this workflow’s runs, so read it as context only.

What to take from it Section titled “What to take from it” Across these three runs, the workflow took between 9 and 23 turns. The lever is how many of those turns go to gathering data before any judgment happens. Whether that gathering can move into a script is the question the audit raises. The next few runs will show whether the November 14 spike was a one-off or a pattern. Try it Section titled “Try it” Read the workflow source to see how the personas are chosen and how each run reports back.

To build something similar, start at github/gh-aw . 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. What the last three runs show Section titled “What the last three runs show” Run Date (UTC) Duration Turns Audit tokens Discussion 43525284919 Nov 13 21.6m 60 3.25M #77583 43651937051 Nov 14 26.5m 3 547k #77876 43773614833 Nov 15 24.1m 2 339k #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. Signals worth a closer look Section titled “Signals worth a closer look” 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. What to take from it Section titled “What to take from it” 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.

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

Release: v1.2.3 Section titled “Release: v1.2.3” 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. Unp

The rest of this page is on the live site.