Detent dropped on GitHub this week with a pitch that should sound familiar to anyone who's tried to run autonomous coding agents at scale: manage the work, not the agents. The tool is status-driven agentic orchestration shipped as a single CGO-free Go binary, and it brings some interesting ideas to how AI-coded PRs actually get merged safely.

How Worktrees Change the Isolation Game

The core unit of work in Detent is an issue on a board—either GitHub Projects v2 or a boardless setup using repository status labels. When you mark something Todo, Detent claims it, creates an isolated Git worktree from your source checkout, and fires off a Codex agent with the workflow contract you wrote. Each agent works in its own branch, runs your validation gate, opens or updates a PR, then moves the issue to Human Review. Using actual worktrees instead of subdirectories or temp clones is clever—it's proper filesystem isolation without the overhead of a full repo copy. If you're running twenty concurrent issues, you get twenty worktrees all chugging away independently. That's not something you can safely do with naive branch-per-issue approaches unless you like debugging merge conflicts that shouldn't exist.

The Serialized Merge Train

Here's where Detent gets opinionated: the merge train is serialized. One rebase, CI watch, and merge at a time. Concurrent candidates never invalidate each other's CI status because they literally cannot run simultaneously in the Merging lane. This is a direct response to the chaos that happens when you let parallel agent PRs all try to land at once—stale branches, race conditions on main, green checks that lied. Gates decide what promotion looks like. The workflow decides whether Merging waits for a human label, an automated PR review against current head, or just CI passing plus quiet time. Unresolved feedback bounces it back to Rework for another pass. It's a proper state machine with explicit transitions—nothing implicit, nothing hidden.

Why Codex Over Claude

Detent dispatches agents through OpenAI's Codex CLI via codex app-server, not Anthropic's Claude Code. The docs spell out the economics: ChatGPT plans (Plus, Pro, Business) cover Codex CLI usage including scripted automation under your existing subscription. Starting June 15, 2026, Anthropic moved headless claude -p, the Agent SDK, and third-party agent apps to a separate monthly credit that is per-user, doesn't roll over, and bills at standard API rates for overages. For an orchestrator running agents around the clock across multiple repositories, those metering differences matter. The source material frames it bluntly: "For an orchestrator that runs agents around the clock in parallel, the Codex subscription is the one that makes the economics work." Fair enough. Detent keeps the backend configurable via agents.backends so alternatives remain viable as automation terms evolve.

One Host, Many Repos

The global.yaml config lets you run multiple projects with weights, priority, pause controls, and fair scheduling from a single Detent host. The web dashboard shows live counts, running agents, token usage, budget state, and board flow—useful when you've got five repositories all feeding into the same pipeline. Detent's own production config at digitaldrywood/detent-orchestration dispatches the agents that build Detent itself. That's a nice bootstrap story: use the tool to build the tool, with proper gates on everything.

Key Takeaways

  • Worktree-based isolation means true parallelism without filesystem pollution or full repo copies
  • Serialized merge train prevents CI invalidation from concurrent landing attempts—a real operational hazard at scale
  • Explicit workflow contracts (WORKFLOW.md) encode your process; Detent executes it with discipline, not judgment
  • Multi-project support with weighted scheduling from one binary—no BEAM runtime or service mesh required

The Bottom Line

Detent is interesting because it's pragmatic where others are ideological. It doesn't try to make agents smarter—it tries to make your engineering process run exactly as specified, every time, in parallel. Whether that contract-first philosophy actually holds up at scale depends on whether your team's workflow definitions are worth trusting. If they are, Detent removes the supervision overhead that's been killing autonomous agent adoption in production.