PATTERN Cited by 1 source
Agentic PR triage¶
Shape¶
Give a coding agent:
- An ephemeral dev environment it owns,
gh(or a similar CLI / API client) authenticated against a GitHub org / repo, and- A filter over an issue tracker (labels, assignees, project boards) that defines the agent's queue of candidate work.
Let the agent pick issues autonomously, do the work in the ephemeral environment, and surface results as pull requests. The human reviews the PRs asynchronously — possibly hours later — rather than sitting through each edit-compile-test cycle.
Canonical statement¶
Chris McCord, on Phoenix.new (2025-06-20):
I'm already using Phoenix.new to triage
phoenix-coreGithub issues and pick problems to solve. I close my laptop, grab a cup of coffee, and wait for a PR to arrive — Phoenix.new knows how PRs work, too.(Source: sources/2025-06-20-flyio-phoenixnew-remote-ai-runtime-for-phoenix)
Why it's a distinct pattern¶
The shape is adjacent to:
- Background IDE agents (Cursor / Copilot background agents) — human still in the editor; agent does long-running tasks inside the session but the session is live.
- Devin-style autonomous engineers — same async shape, packaged as a standalone agent product rather than a "cloud IDE the human signs into".
- Dependabot / Renovate / automated-PR bots — generate PRs
for a narrow class of mechanical work (dep bumps, security
patches). Agentic PR triage generalises this to "any work the
agent can complete with Linux +
gh+ its toolchain".
What makes it a wiki-worthy pattern:
- It's the outer loop of agentic development. The agentic development loop is the inner cycle (agent ↔ execution environment inside a session); agentic PR triage is the cycle over sessions, driven by a work queue and terminated by a PR.
- It productises the async agent workflow concept specifically against GitHub's issue / PR primitives.
- It's the coding-agent specialisation of the 2025-04-08 RX thesis.
Implementation ingredients¶
- Ephemeral dev environment per task. Phoenix.new uses a fresh Fly Machine; Codespaces / Gitpod / Replit could play the same role. See patterns/ephemeral-vm-as-cloud-ide.
- Work-queue shape. GitHub issues with label filters, project
board columns, a
triage-melabel, etc. The agent needs a structured query it can poll. - PR-authoring toolchain. The agent needs to
gh auth, create branches, push, open PRs, write PR bodies, link issues.ghcovers all of this; a narrower agent-specific API (or an MCP wrapper) would do too. - Review gate. A human reviews the PR before merge. Branch protection + CODEOWNERS rules are the load-bearing safety net.
- Feedback loop on PRs. An agent that can read its own PR's CI status, address review comments, and push new commits in response is a materially better experience. The 2025-06-20 post says "Phoenix.new knows how PRs work, too" at this level of resolution.
Trade-offs¶
- Pro: async human review; human free to close laptop; agent productivity scales with wall-clock time rather than human supervision time; natural human-in-the-loop at merge.
- Con: cost scales with wall-clock time, not work completed (an agent looping unproductively for 8 hours bills 8 hours); trust boundary sits at repo-level controls (branch protection, CODEOWNERS) rather than per-action permissions; agent misbehaviour surfaces only at PR-review time.
Caveats¶
- Trust scope.
ghauth withreposcope gives the agent write access to every repo the operator can touch. Per-repo fine-grained PATs, GitHub-App installation tokens, or short-lived OIDC are safer; Phoenix.new doesn't disclose what scope itsghinstall uses. - Work-picking quality. An agent that picks issues outside
its competence burns time. Real deployments will want issue-
level labels the agent respects (
ai-ok,ai-ignore). - Repository-level guardrails are load-bearing. Without
branch protection,
CODEOWNERS, and GitHub Actions limits on AI-generated PRs, the pattern is dangerous. The 2026-03-09 Datadog hackerbot-claw retrospective is the canonical counter-example: attacker agents opening malicious PRs against public repos were contained because the repos had org-wide rulesets. - Cost-per-PR economics are invisible in the 2025-06-20 post. No disclosure of how much an agent-written-and-merged PR costs in session-VM time + LLM tokens.
- Review-fatigue risk. Lots of small agent-generated PRs can overwhelm a reviewer; same UX failure mode as Dependabot overload.
Adjacent patterns¶
- patterns/ephemeral-vm-as-cloud-ide — the substrate that gives the agent per-session VMs to work in.
- patterns/disposable-vm-for-agentic-loop — inner-loop substrate.
- patterns/llm-pr-code-review (Datadog BewAIre, 2026-03-09)
— same
gh-on-PRs shape used defensively against the attacker-agent variant of this pattern. - patterns/specialized-reviewer-agents (Cloudflare AI code review, 2026-04-20) — the reviewer side; an agentic-PR- triage pipeline benefits from the reviewer side being agentic too.
Seen in¶
- sources/2025-06-20-flyio-phoenixnew-remote-ai-runtime-for-phoenix
— canonical statement. McCord uses Phoenix.new +
ghagainstphoenix-coreissues. - sources/2025-04-08-flyio-our-best-customers-are-now-robots — the RX thesis this pattern operationalises for coding agents.
Related¶
- concepts/async-agent-workflow — parent concept.
- concepts/agentic-development-loop — inner loop.
- concepts/robot-experience-rx — grandparent framing.
- systems/phoenix-new — canonical enabling platform.
- systems/gh-cli — the GitHub CLI that makes it executable.
- systems/github — the platform.
- patterns/ephemeral-vm-as-cloud-ide — substrate pattern.