Build one feature with parallel coding agents
Write the feature as one task. Your own coding agent plans it and splits it into pieces that never touch the same files. Each piece is built by an agent in its own sandbox, at the same time. TaskWing joins the work on one branch, runs a hidden exam in a clean sandbox and opens one draft pull request. In our largest run so far, 2 agents wrote 2,076 lines with 0 merge conflicts.
When this fits
A feature that crosses several parts of a codebase: an API route, a data model, a UI screen, a background job, the docs. One agent working through it in order is slow, and several agents in one folder step on each other. TaskWing gives each part to its own agent, in its own sandbox, and keeps them apart by file ownership.
It fits less well when the change is one tight edit in one file, or when you want to steer every step yourself. For those, run your coding agent directly.
How a run goes
- You give one task. For example:
taskwing run "add discount codes at checkout, with tests", from a clean working tree on the branch the pull request should target. - The lead plans. Your coding agent, on your machine, splits the task into pieces and gives each one the paths it owns. It also pins the interface between them, so two agents agree on names before they start.
- A separate session writes the exam. Plain-sentence checks plus hidden tests. You read the plan and the checks and press
y. No worker ever sees the hidden tests. - Agents build in parallel. One sandbox and one agent per piece. Code goes in as a git bundle; only commits come out. The terminal shows one live square per agent.
- The work is joined and examined. Every piece's commits land on one branch, and the exam runs in a clean sandbox that no worker touched.
- One draft pull request. Pushed from your machine, with the exam results and every agent's log saved under
~/.taskwing/runs/.
What a real run looked like
We pointed TaskWing at Osprey, an open source transaction monitoring engine in Go, with one task: add an alert webhook, an endpoint to list evaluations and a disposition API.
| Measure | Result |
|---|---|
| Pieces and agents | 2 |
| Lines written | 2,076 |
| Merge conflicts at the join | 0 |
| Time | 15 min 12 s |
| Hidden exam | 3 of 4 checks passed |
The failed check was our fault, not the agents': the exam writer guessed a JSON field name that the plan never fixed. We changed TaskWing so the lead pins the interface first and the exam is written from it. The full story is in why the agents never see the tests.
What you need
- A git repository with a clean working tree, plus
git, theghCLI and Claude Code, all logged in on your machine. - A boat account and a base box with your repository's toolchain. The first base took about an hour by hand; a
taskwing setupcommand is planned. - Or, to see the flow first,
--provider local: agents work in folders on your machine with no isolation.
Limits today
- Proof of concept, command line only. The desktop app draws runs from a demo engine until the engine stream ships.
- One agent runtime (Claude Code) and one sandbox provider (boat).
- No resume: a crashed run starts again.
taskwing doctor --cleanremoves leftover sandboxes.
Questions
How many agents work on one feature?
As many as the plan has pieces, up to a cap you set with --max-cards (4 by default). TaskWing never promises a fixed number: a small task may be one piece.
What stops two agents from editing the same file?
The plan check. Before any agent starts, TaskWing rejects a plan in which two pieces own the same path, and the lead names the shared interface (routes, fields, environment variables) up front.
What happens if pieces conflict when they are joined?
Today a piece that conflicts at the join is dropped from the run, and the pull request says so. Having the lead resolve conflicts is planned for the next milestone.
Does TaskWing merge the feature?
No. A run ends at one draft pull request with the exam results. You review it and you merge.
Which agent and model build the pieces?
Claude Code today, with the model and effort you pass in. A second agent runtime, Codex CLI or OpenCode, is the next one planned.
Try the proof of concept
TaskWing is open source (AGPL-3.0) and early. Today it runs Claude Code agents in boat sandboxes, or in local folders to try the flow.
go install github.com/josephgoksu/TaskWing/cmd/taskwing@main
taskwing run "add a /health endpoint with a test" --provider local --no-pr