The Tool Renamed Itself

Before announcing localagent anywhere, I wanted to know the neighborhood. So the orchestrator did what orchestrators do: fanned out three research agents, each sweeping a different angle — established coding CLIs, worktree-isolation tools, and orchestrator/worker patterns.

The findings compressed to four lines:

1. Local-model support is table stakes. Every surviving coding agent speaks Ollama or LM Studio. That alone differentiates nothing.
2. Worktree isolation is commoditized. By mid-2026 nearly every serious tool ships it; Claude Code has it as a first-party feature.
3. Nobody with traction ships the combination — a local model running a real tool loop in an isolated worktree, leaving a full JSONL transcript that a stronger frontier model audits before anything merges. Aider's architect/editor mode hands work down but never audits it afterward. Qwen Code ships worktrees and a review command — the same model grading its own homework. The dozen tiny MCP servers with the "Claude thinks, your local model grunts" tagline forward single prompts and get text back — no tool loop, no transcript, nothing to audit.
4. The ecosystem's big bet is replacement. Ollama and LM Studio both shipped Anthropic-compatible endpoints in January 2026 so that Claude Code can run entirely on local models. Swap the frontier model out; keep the workflow.

Our bet is the inverse of point four. Keep the frontier model — but only where judgment lives, as planner and auditor. Hand everything else to the models already running for free on the Mac in the corner.

Did "localagent" say any of that? It did not. It was a repo called localagent, a command called localagent, and a PyPI dist called dirtsimple-agent — every obvious dist name had tripped PyPI's similarity filter — two names spread across three places, none of them saying anything at all.

dirtywork. As in: frontier models do the thinking, local models do the dirty work. The whole architecture in one phrase.

The domain hunt was brief. dirty.work — taken, confirmed against the registry's RDAP records. dirtywork.com — a premium listing at $57,500, which is a lot for a tool whose entire pitch is not paying for things. dirtywork.run — available, $6.99 for the first year. Bought for three years on the spot.

And the domain is an imperative. dirtywork, run. It's also literally the command.

A rename touches everything — package, CLI, launcher, tests, packaging metadata, README, blog, DNS, PyPI trusted publishing. So it got the same treatment any feature gets here: a spec, an eleven-task implementation plan, and subagent-driven execution with a reviewer gating every task and a held PR at the end.

Two of those eleven tasks were the mechanical heart of the job: rename the Python package (imports, branch prefixes, test assertions, launcher) and sweep the docs. Bounded, well-specified, verifiable — exactly the shape of task the tool exists for, and exactly the shape of task I'd normally delegate to it against someone else's repo. So the tool got assigned its own rename. A qwen3-coder worker, in an isolated worktree, renaming the harness it was running inside of.

Task one went cleanly, and in the right order: the worker updated the test expectations first, ran the suite to watch it fail with ModuleNotFoundError: No module named 'dirtywork' — red for the right reason — then renamed the package and brought all 133 tests back to green. 24 turns. The task review approved it with nothing to fix.

Then task three — the docs sweep — refused to start. Exit code 127: command not found.

The localagent command on my PATH was a symlink to bin/localagent in the repo. Task one had just renamed that file to bin/dirtywork. The launcher I was using to launch the rename had been renamed out from under me by the rename. The first casualty of the rename was the renamer.

One relaunch through the new launcher later, the worker ran fine. And the run's own receipt is the part I'd frame: it came back reporting a worktree at .worktrees/dw-…, on a branch named dirtywork/…, with its transcript under ~/.dirtywork/runs/. Every one of those names was minutes old. Task one hadn't even merged — the worker was running inside the renamed harness on its own feature branch, which means the rename was validated in production by the rename itself, before its own pull request existed.

That run: 47 turns, about 1.1 million prompt tokens, $0 in API fees.

Here's what didn't go clean.

The worker missed one plan line item — a couple of README examples still showed the old worktree prefix. Half of that is on the orchestrator: the line existed in the plan and didn't survive into the worker's prompt. Specs don't delegate themselves.

More interesting: the worker added things. Twice, in one docs task, it invented decorative HTML nobody asked for — a styled stats box here, an extra flourish there. Reviewers caught both. That's the sharpened lesson of the day: when you audit a local worker's diff, you're not just looking for what's missing. You're looking for what's extra.

And the final whole-branch review caught a defect that wasn't the worker's at all: the module's __version__ still said 0.1.0 because the plan never listed that file. The review gate doesn't just catch the cheap model. It catches the orchestrator writing the plan. That's the point of gates.

A rename is the purest kind of release: zero functional changes intended, which means any behavior difference at all is a bug. That makes it a clean test of the whole harness — and the harness passed it on itself.

• Competitive survey → name → domain bought: about half an hour
• Spec, an eleven-task plan, and a six-commit PR (two commits carrying the tool's own work), one worker fix round, one final-review fix
• Repo renamed (old URLs redirect), DNS live with the Pages IPs checked against GitHub's own docs, TLS issued, HTTPS enforced
• dirtywork 0.2.0 on PyPI: pipx install dirtywork

The first post ended by saying the pattern needed more reps before I'd call it proven. It got its reps the same evening, on itself: same worktrees, same transcripts, same review gate — pointed inward. The tool that delegates dirty work spent the night doing its own, exactly as drawn.

Frontier models do the thinking. Local models do the dirty work. Now the name says so.