Put a coding agent in your pipeline.

A coding agent in CI is an AI agent that runs inside your GitHub Actions workflows. It reviews pull requests, makes fixes when someone asks in a comment, and triages new issues. It leaves its work as comments, commits or a pull request, and a human still decides what merges.

As of September 2026, GitHub, Anthropic, OpenAI and Google all offer an official way to do this. This guide compares them and walks through a minimal workflow file copied from Anthropic’s docs. Then it covers the part that matters most: what the agent can touch, who can trigger it, and what it costs.

01What a coding agent can do in CI

Four jobs cover most of the value:

• Review pull requests. The agent reads the diff when a PR opens and comments on bugs, risks and missing tests.

• Fix on request. A teammate comments “@claude fix the failing test in checkout”, and the agent pushes a commit to the branch.

• Triage issues. It labels new issues, spots duplicates and asks the reporter for missing steps to reproduce.

• Run scheduled chores. A nightly job summarizes yesterday’s commits or open issues.

Reviews are the safest place to start, because the agent only comments. If reviews are all you need, compare dedicated AI code review tools before you build a workflow.

02The official options in 2026

There are three shapes. GitHub runs the agent for you, a vendor Action runs inside a workflow you own, or GitHub compiles a plain-language workflow into a locked-down Action.

How to choose: if you want no YAML at all, assign issues to Copilot, Claude or Codex on GitHub. If you want control over triggers, prompts, models and permissions, use a vendor Action. Agentic Workflows suit many small automations under strict guardrails: they are read-only by default, and writes go through validated “safe outputs”. They can drive Copilot, Claude Code, Codex, Gemini or Pi.

Still choosing an agent for day-to-day work? Our Claude Code, Codex and Cursor comparison covers how they differ outside CI.

One caveat for Gemini: in June 2026, Google moved consumer users from Gemini CLI to its new Antigravity CLI. Gemini CLI, which the Gemini Action runs, stays available with paid Gemini API keys and enterprise licenses.

03Add a coding agent to GitHub Actions in one file

This is the minimal setup from Anthropic’s documentation for the Claude Code GitHub Action. The Codex and Gemini Actions follow the same pattern: an API key in Secrets, a workflow file and a trigger.

1. 01

   Install the GitHub App

   Install the Claude GitHub App on the repository. You need admin access. Running /install-github-app inside Claude Code installs the app and adds the secret for you, then prepares a pull request with the workflow.

2. 02

   Add the key as a secret

   In the repository settings, open Secrets and variables, then Actions, and add ANTHROPIC_API_KEY. Subscribers can use a CLAUDE_CODE_OAUTH_TOKEN from claude setup-token instead.

3. 03

   Commit the workflow

   Save the file below as .github/workflows/claude.yml on your default branch.

4. 04

   Mention the agent

   In any issue or PR comment, write @claude followed by the request. Claude replies in a comment on the same thread and updates it as it works.

name: Claude Code
on:
  issue_comment:
    types: [created]
  pull_request_review_comment:
    types: [created]
jobs:
  claude:
    if: contains(github.event.comment.body, '@claude')
    runs-on: ubuntu-latest
    permissions:
      contents: write
      pull-requests: write
      issues: write
      id-token: write
      actions: read
    steps:
      - uses: actions/checkout@v6
        with:
          fetch-depth: 1
      - uses: anthropics/claude-code-action@v1
        with:
          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

The if line stops the job before a runner starts on comments that do not mention @claude. The permissions block lets the agent push commits and comment, id-token: write serves the Action’s GitHub App sign-in, and actions: read lets it read CI results.

We copied this file from the docs as of September 2026. Both actionlint 1.7.12 and a check against the public JSON schema for GitHub workflows passed it. We could not run it on a live repository, because that step needs your app install and your API key.

04Lock down permissions and secrets

An agent in CI holds real credentials and reads text written by strangers. Plan for both:

• Grant the least access. A review-only job does not need write access to code. Anthropic’s own review example uses read permissions for contents, pull requests and issues.

• Keep keys in Secrets. Never write a key into the workflow file. Use a dedicated CI key you can revoke on its own, and follow our guide to keeping API keys safe.

• Limit who can trigger it. The Claude Action requires write access from whoever triggers it on issue and PR events, and rejects bots unless you allow them. Copilot’s agent also answers only to people with write access.

• Treat all text as untrusted. Issue bodies, comments and diffs can carry instructions aimed at the agent. Allow only the tools a job needs (--allowedTools in claude_args), and review everything it writes.

• Leave fork pull requests alone. On public repositories, GitHub withholds secrets from runs triggered by fork PRs. Do not work around that with pull_request_target and a checkout of the fork’s code, which GitHub’s security guide warns against.

• Pin the Action. GitHub calls pinning to a full commit SHA “the only way to use an action as an immutable release”.

GitHub’s hosted agents add their own limits. Copilot’s cloud agent pushes to a single branch, usually a new copilot/ one, and workflows on its pull requests wait until someone with write access approves them. The person who asked for its pull request cannot approve it, and its internet access runs through a firewall. Claude and Codex on GitHub get the same protections.

The agent’s changes still need the same checks as any AI-written code. Our checklist for securing AI-generated code covers the scanners to run on its pull requests.

05How much does a coding agent in CI cost?

Every run has two meters. Runner time counts against your GitHub Actions minutes, and the model bills tokens to your API key or subscription. Copilot’s agent, and Claude and Codex on GitHub, use Actions minutes plus AI credits from your Copilot plan.

Cap the two obvious leaks: runs nobody needed and runs that never end. These additions to the workflow above do both, and we validated the combined file the same way:

    timeout-minutes: 15
    concurrency:
      group: claude-${{ github.event.issue.number || github.event.pull_request.number }}

          claude_args: "--max-turns 10"

The first snippet goes under runs-on. Its timeout ends a stuck job, and concurrency allows one run per issue or pull request at a time. The second goes under anthropic_api_key and limits how many turns the agent takes. Beyond that, write specific requests, keep CLAUDE.md short because the agent reads it on every run, and use a cheaper model for routine reviews with --model.

06FAQ

Read next: how to review AI-written code when the agent’s pull request lands.