Research
What the numbers say.
Who actually runs coding agents from GitHub Actions, what each agent supports, and whether skills help. Every number links to its source. Last checked Oct 17, 2026.
- repos run agents from YAML
- ~11k
- repos run agents from YAML
- pts from curated skills
- +16.6
- pts from curated skills
- skills beat big bundles
- ≤3
- skills beat big bundles
Adoption
YAML-driven agents are real, and still early.
The only direct count we found searched public .github/workflows files for agent actions and found 13,392 workflows in 10,792 repos. Lower bounds from GitHub code search. The paper does not give its collection date, and Cursor and gh-aw are not counted.
Workflows per action
- anthropics/claude-code-action5,212
- actions/ai-inference795
- openai/codex-action492
- testdriverai/action300
- google-gha/run-gemini-cli244
- anthropics/claude-code-base-action227
- austenstone/copilot-cli195
Copilot is an exact bot-account count. The others are estimated from declared attribution and include local runs, not only CI.
Support
Which agent does what.
I need:
Pick what you need to see which agents fit.
| Capability | claude-code-action | codex-action | Cursor | gh-aw | Copilot agent |
|---|---|---|---|---|---|
| Runs from your workflow YAMLCan you add it as a step in your own .github/workflows file? | Built inuses: anthropics/claude-code-action | Built inuses: openai/codex-action | Built inInstall the CLI or curl the Cloud Agents API | Built inMarkdown source compiles to a .lock.yml workflow | Not supportedGitHub runs it for you; there is no uses: step |
| @mention in a commentSomeone types @agent in an issue or PR and the agent answers. | Built intrigger_phrase, default @claude | Write it yourselfWrite the if: contains(...) check yourself | Write it yourselfWrite the trigger yourself | Built inslash_command | Built inComment-triggered automations (Aug 2026) |
| Label triggerAdding a label such as ai-review starts the agent. | Built inlabel_trigger | Write it yourselfon: issues labeled + an if: check | Write it yourselfon: issues labeled + an if: check | Built inlabel_command, removes the label after | Not supportedNot documented |
| Assign an issueAssigning the issue to the agent hands it the work. | Built inassignee_trigger | Write it yourselfon: issues assigned | Write it yourselfon: issues assigned | Write it yourselfon: issues assigned | Built inThe main way to start it |
| ScheduleRuns on a timer, for example a nightly maintenance pass. | Built incron + automation mode | Write it yourselfcron | Write it yourselfcron | Built inFuzzy schedules such as "daily" | Not supportedNot documented |
| Fix failing CIWakes up when another workflow fails and tries a fix. | Write it yourselfworkflow_run + actions: read | Write it yourselfworkflow_run | Built inCookbook has a fix-CI example | Built inSee the CI coach workflow | Not supportedNot documented |
| Structured outputThe agent returns JSON that matches a schema, so later steps can use it safely. | Built in--json-schema | Built inoutput-schema | PartialJSON output, no schema enforcement | Built inTyped safe-outputs | Not supportedNot applicable |
| MCP toolsThe agent can call Model Context Protocol servers during the run. | Built in--mcp-config; GitHub MCP built in | PartialConfigure through codex-home | Built inmcpServers in the API | Built intools: and imports: | Built inRepository MCP settings |
| Pay with a subscriptionUse a plan you already pay for instead of a metered API key. | Built inCLAUDE_CODE_OAUTH_TOKEN (Pro, Max, Team, Enterprise) | Not supportedOpenAI API key or Azure only | Built inBilled to your Cursor plan usage | PartialCopilot engine yes; ignores CLAUDE_CODE_OAUTH_TOKEN | Built inCopilot seat, AI Credits since June 2026 |
Checked against vendor docs on Oct 17, 2026: Claude Code GitHub Actions docs · openai/codex-action README · Cursor CLI in GitHub Actions · gh-aw frontmatter reference · gh-aw auth reference · About Copilot coding agent · Copilot automations from comments
Skills
Do skills actually make agents better?
+16.6 pts
Curated skills help on average
Across 87 tasks and 18 model setups, human-written skills raised pass rates from 33.9% to 50.5%. Software engineering gained less: +11.6 points.
SkillsBench, arXiv 2602.12670−8 to −11.5
Self-written skills hurt
Skills an agent wrote for itself scored below having no skill at all. Write or curate them by hand.
SkillsBench, arXiv 2602.12670≤3 modules
Focused beats exhaustive
Small skills beat big bundles. Skills that forced one rigid pipeline made results worse on 13 of 87 tasks.
SkillsBench, arXiv 2602.12670+20% cost
AGENTS.md rarely lifts success
Context files generally did not raise task success and added over 20% inference cost. They helped most for unusual conventions.
Evaluating AGENTS.md, arXiv 2602.11988−29% time
But it can make runs faster
A second study saw median runtime drop about 29% and output tokens about 17%, with similar completion.
AGENTS.md efficiency, arXiv 2601.20404
How we rate each skill
- Research-backedPublished studies support the core practice.
- Practitioner-backedIndustry practice and surveys support it; controlled studies are thin or mixed.
- OpinionPlausible craft advice that nobody has measured.
| Skill | Rating | Why |
|---|---|---|
| Code review | Practitioner-backed | Code review itself is well studied: low review coverage predicts more post-release defects. The checklist format this skill uses is not; a 2022 study found no strong link between checklists and review performance. LLM reviewers still miss most issues (best F1 about 19% on SWR-Bench), and structured prompts help mainly by cutting false positives. |
| Debug | Research-backed | Observed professionals reproduce, build a model, fix, and validate, and they lean on recent changes, which is the loop this skill teaches. Offering developers candidate hypotheses made them 6x more likely to succeed. The agent-side gain is unproven: LLM "scientific debugging" did not consistently beat plain repair. |
| Deploy checklist | Practitioner-backed | DORA research supports the canary, monitoring, and rollback parts. It also finds heavyweight external change approval adds no stability benefit, which cuts against this skill's approvals items. Nothing measures agents writing deploy checklists. |
| Documentation | Practitioner-backed | DORA links documentation quality to organizational performance, and 93% of open-source contributors call outdated docs a pervasive problem. The templates themselves are convention, and nothing measures a docs skill for agents. |
| Frontend design | Opinion | Well-crafted taste from the vendor, but untested: the published write-up shows before-and-after screenshots with no metrics or blind comparison. Critique-then-revise loops have some support when they get real feedback such as screenshots. |
| System design | Practitioner-backed | The requirements-first framework mirrors textbook practice. Clarifying ambiguous requirements before coding raised GPT-4 pass rates from 68.0% to 75.8% in one study, though that measured code generation, not design. |
| Testing strategy | Practitioner-backed | The testing pyramid's proportions are convention: one study found unit and integration tests catch the same defect types. Handing an LLM the tests up front does raise its pass rates, which supports test-first agent loops. |
| Web app testing | Practitioner-backed | Running bundled scripts instead of loading them into context is a reasonable design argument without a published evaluation. Focused skills like this one beat exhaustive bundles on SkillsBench. One instruction contradicts Playwright's own docs (see the caveat). |
| Browser testing with Chrome DevTools MCP | Evidence not reviewed | Not reviewed yet. |
| CI/CD and automation: set up build and deploy pipelines | Evidence not reviewed | Not reviewed yet. |
| Multi-axis code review before merge | Evidence not reviewed | Not reviewed yet. |
| Constraint-driven development: stop agents quietly lowering the quality bar | Evidence not reviewed | Not reviewed yet. |
| Context engineering: set up agent context for quality output | Evidence not reviewed | Not reviewed yet. |
| Debugging and error recovery: systematic root cause | Evidence not reviewed | Not reviewed yet. |
| Doubt-driven development: fresh-context adversarial review of decisions | Evidence not reviewed | Not reviewed yet. |
| Git workflow and versioning: commits, branches, releases | Evidence not reviewed | Not reviewed yet. |
| Incremental implementation: thin, verifiable slices | Evidence not reviewed | Not reviewed yet. |
| Planning and task breakdown: spec to ordered tasks | Evidence not reviewed | Not reviewed yet. |
| Security and hardening: audit inputs, auth and secrets | Evidence not reviewed | Not reviewed yet. |
| Source-driven development: ground decisions in official docs | Evidence not reviewed | Not reviewed yet. |
| Spec-driven development: write the spec before code | Evidence not reviewed | Not reviewed yet. |
| Test-driven development: red-green-refactor loop | Evidence not reviewed | Not reviewed yet. |
| Agent harness construction: action spaces, tools, observations | Evidence not reviewed | Not reviewed yet. |
| Agentic engineering: eval-first, decomposed, cost-aware execution | Evidence not reviewed | Not reviewed yet. |
| AI regression testing for agent-written code | Evidence not reviewed | Not reviewed yet. |
| Autonomous loops: patterns for unattended Claude Code pipelines | Evidence not reviewed | Not reviewed yet. |
| Context budget: check live context use in Codex or Claude Code | Evidence not reviewed | Not reviewed yet. |
| Android skills: the android-cli entry skill from the official Android pack | Evidence not reviewed | Not reviewed yet. |
| Subagent development: create and tune Claude Code subagents | Evidence not reviewed | Not reviewed yet. |
| Build an MCP server: end-to-end server scaffolding guide | Evidence not reviewed | Not reviewed yet. |
| Claude API reference: models, params, streaming, tool use | Evidence not reviewed | Not reviewed yet. |
| Claude Code automation recommender: suggests hooks, subagents, skills, MCP | Evidence not reviewed | Not reviewed yet. |
| CLAUDE.md improver: audit and tighten your CLAUDE.md | Evidence not reviewed | Not reviewed yet. |
| Slash command development: build custom commands | Evidence not reviewed | Not reviewed yet. |
| Hook development: write PreToolUse/PostToolUse/Stop hooks | Evidence not reviewed | Not reviewed yet. |
| MCP builder: guide to building high-quality MCP servers | Evidence not reviewed | Not reviewed yet. |
| MCP integration: add and configure MCP servers in a plugin | Evidence not reviewed | Not reviewed yet. |
| Session report: HTML report of tokens, cache, subagents and skills usage | Evidence not reviewed | Not reviewed yet. |
| Skill creator: write, improve and eval new skills | Evidence not reviewed | Not reviewed yet. |
| Skill development: author skills for Claude Code plugins | Evidence not reviewed | Not reviewed yet. |
| Hookify rules: write guardrail rules without hook code | Evidence not reviewed | Not reviewed yet. |
| Caveman: short-output skill that cuts agent tokens | Evidence not reviewed | Not reviewed yet. |
| Cloudflare Agents SDK: build, debug and review agents apps | Evidence not reviewed | Not reviewed yet. |
| Cloudflare Durable Objects: stateful coordination code | Evidence not reviewed | Not reviewed yet. |
| Web performance audit: Core Web Vitals and Lighthouse | Evidence not reviewed | Not reviewed yet. |
| Cloudflare Workers best practices for production | Evidence not reviewed | Not reviewed yet. |
| Wrangler: run and troubleshoot Cloudflare Wrangler CLI | Evidence not reviewed | Not reviewed yet. |
| EAS workflows: write Expo CI/CD workflow YAML | Evidence not reviewed | Not reviewed yet. |
| Expo SDK upgrade and dependency fixes | Evidence not reviewed | Not reviewed yet. |
| Claude settings audit: recommend settings.json permissions | Evidence not reviewed | Not reviewed yet. |
| Commit: conventional commit messages (Sentry style) | Evidence not reviewed | Not reviewed yet. |
| Find bugs: bugs, vulns and quality issues in branch changes | Evidence not reviewed | Not reviewed yet. |
| GitHub Actions security review: find workflow exploitation bugs | Evidence not reviewed | Not reviewed yet. |
| Iterate on a PR until CI passes and review feedback is addressed | Evidence not reviewed | Not reviewed yet. |
| PR writer: reviewer-facing PR titles and descriptions | Evidence not reviewed | Not reviewed yet. |
| Skill scanner: scan agent skills for security issues | Evidence not reviewed | Not reviewed yet. |
| Skill writer: create and improve skills per the Agent Skills spec | Evidence not reviewed | Not reviewed yet. |
| gh-aw agentic workflows: design, create, debug and upgrade | Evidence not reviewed | Not reviewed yet. |
| Checkout credential review: safe git/gh ops in agentic workflows | Evidence not reviewed | Not reviewed yet. |
| Debug gh-aw workflow failures from logs and audits | Evidence not reviewed | Not reviewed yet. |
| Optimize agentic workflow token usage (gh-aw) | Evidence not reviewed | Not reviewed yet. |
| Review agentic workflow changes for correctness and security | Evidence not reviewed | Not reviewed yet. |
| gstack: Garry Tan's Claude Code skill set | Evidence not reviewed | Not reviewed yet. |
| Hugging Face Hub CLI: models, datasets and Spaces | Evidence not reviewed | Not reviewed yet. |
| Hugging Face model evals with inspect-ai and lighteval | Evidence not reviewed | Not reviewed yet. |
| Code review (Matt Pocock): review changes since a fixed point on two axes | Evidence not reviewed | Not reviewed yet. |
| Diagnosing bugs: loop for hard bugs and perf regressions | Evidence not reviewed | Not reviewed yet. |
| Grill with docs: interview to sharpen a plan and write ADRs | Evidence not reviewed | Not reviewed yet. |
| Improve codebase architecture: find deep-module opportunities | Evidence not reviewed | Not reviewed yet. |
| PR body writer (Matt Pocock) | Evidence not reviewed | Not reviewed yet. |
| TDD (Matt Pocock): build features and fix bugs test-first | Evidence not reviewed | Not reviewed yet. |
| To tickets: turn a plan into tracer-bullet tickets | Evidence not reviewed | Not reviewed yet. |
| Triage: move issues and PRs through a triage state machine | Evidence not reviewed | Not reviewed yet. |
| Wayfinder: plan work bigger than one agent session | Evidence not reviewed | Not reviewed yet. |
| Superpowers: plan, test, and review skills framework | Evidence not reviewed | Not reviewed yet. |
| Playwright skill: browser-test your app from the agent | Evidence not reviewed | Not reviewed yet. |
| pstack for Claude Code, Codex, and more | Evidence not reviewed | Not reviewed yet. |
| Stripe integration best practices | Evidence not reviewed | Not reviewed yet. |
| Upgrade Stripe API versions, SDKs and webhooks | Evidence not reviewed | Not reviewed yet. |
| Supabase: database, auth, storage and edge functions tasks | Evidence not reviewed | Not reviewed yet. |
| Postgres best practices (Supabase) | Evidence not reviewed | Not reviewed yet. |
| SVG diagram skill for architecture and flow charts | Evidence not reviewed | Not reviewed yet. |
| React composition patterns that scale (Vercel) | Evidence not reviewed | Not reviewed yet. |
| React and Next.js performance best practices (Vercel) | Evidence not reviewed | Not reviewed yet. |