Niches · Agent harnesses

Run several agents at once in separate worktrees or hand tasks between harnesses

Four or more companies' docs show a full fit.

10companies, full fit
23companies, fit with a limit
77companies carry the tag
4share of the category's work
2first-hand accounts

What counts as done

Each task ends on its own branch with a reviewable diff, no two agents touched the same checkout, and the orchestrator can tell which finished and which stalled.

How agents ask for it: run three tasks in parallel on separate branches, give this to Codex and that to Claude Code, start a worker for each issue

Does the job 10

The docs show it meets the test above, and nothing we read says otherwise.

ListingWhy
Claude Code C
Anthropic, PBC
claude --worktree <name> starts each session in its own git worktree under .claude/worktrees, and claude agents --json lists background sessions for a script. source
Cursor CLI E
Anysphere, Inc.
-w, --worktree [name] runs each task in a new worktree under ~/.cursor/worktrees, --worktree-base picks the base, and -p runs exit on their own. source
Devin C
Cognition AI, Inc.
devin_session_create starts one or more sessions, each in its own VM, and devin_session_gather tells which settled, errored or are waiting. source
Droid CLI C
The San Francisco AI Factory, Inc.
droid exec --worktree <name> runs each job in its own worktree, the docs show two jobs fanned out in parallel, and clean worktrees are removed on exit while dirty ones stay for review. source
Emdash C
General Action, Inc.
Add Task creates one or more worktrees, each running its own agent from 34 supported CLIs, with diff review, pull requests and CI checks in the app, locally or over SSH. source
Paperclip C
Paperclip Labs, Inc.
Isolated execution workspaces give each task a new git worktree on its own branch, several agents work at once, and task status shows which finished. source
Paseo B
Mohamed Boudra
`paseo run --new-workspace worktree --worktree-mode branch-off --new-branch <name>` gives each agent its own worktree, `paseo ls` lists running agents, and agents can launch other agents across providers. source
Sculptor C
Imbue, Inc.
Each workspace is a git worktree on its own branch, several agents and workspaces run at once, and the app shows changes and a pull request button per workspace. source
T3 Code C
T3 Tools Inc.
Each thread can work in its own worktree, outside agents launch and wait on threads over MCP, and Git actions commit, push and open a pull request. source
Traycer C
Traycer AI
`traycer worktree create --workspace --branch` and the app give each agent its own worktree and branch, agents message each other across tasks, and notifications say which agent finished or failed. source

Does the job, with a limit 23

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

ListingWhy
Amp C
Amp Frontier Corporation
amp -ox or --executor runner:<id> starts a thread elsewhere and prints its URL, so tasks can fan out to orbs or runners, but there is no worktree flag for parallel local runs. source
Clawboo B
Sanreds
Each task-runner task gets its own git worktree, the Board shows running, blocked and Needs you, and a task can pass between Claude Code, Codex, Hermes and OpenClaw, but every child runs with permissions bypassed and no sandbox. source
CLI Manager BB
solhun (Jeong Kyeonghun)
One-click worktree workspaces each with their own sessions, and the API reports idle, awaiting_input and exited per session, but worktree creation is UI only and Git functions are not exposed through the API. source
Cline C
Cline Bot Inc.
--worktree creates a detached worktree per task and -z runs sessions in the background hub, but every child runs with auto-approval on and there is no sandbox. source
Conductor E
Conductor
Isolated workspaces on git worktrees run several agents in parallel with review and merge, but nobody new can download it and existing users lose it in early 2027. source
DeepSeek Harness (dsh) D
DeepSeek (the LICENSE names DeepSeek, and no legal entity is named in the repository or the docs)
The headless README documents running from a worktree with --add-dir granting the shared .git directory, but dsh creates no worktrees itself. source
dsh-superagent (DeepSeek) E
DeepSeek (the LICENSE names the DeepSeek Harness Team; DeepSeek Harness's terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)
Parallel Executor Sessions each in a worktree under a concurrency limit, each ending in its own pull request, but every Executor runs without approval as the operator's user. source
Gemini CLI B
Google LLC
Git worktrees give each session its own copy of the codebase, but the feature is experimental and must be enabled in settings. source
GitHub Copilot CLI C
GitHub, Inc.
--worktree runs or resumes a session in a worktree and --fleet fans out subagents, but exit codes are undocumented so an orchestrator reads the JSON result record instead. source
Harness B
Autonomous, Inc.
A worktree per harness, harness new for Claude Code or Codex and harness wait telling idle, needs the person or stopped, but it needs a browser sign-in and an experimental setting, and Codex runs Harness commands outside its sandbox. source
HarnessRouter D
Lumentree Corporation
Codex, Claude Code, Hermes and Pi run behind one contract and a task moves between them by changing the ID, each in its own sandbox, but there are no git branches or diffs. source
Kilo Code CLI BB
Kilo Code Inc. (the terms are a contract with Anaconda, Inc. on behalf of itself and its affiliates, including Kilo Code Inc.)
--worktree creates or reuses a named worktree and the agent_manager tool starts worktree sessions, but --auto approves every request not denied and the sandbox is off by default. source
Mistral Vibe CLI B
Mistral AI (RCS Paris 952 418 325)
`--worktree <name>` gives each run its own checkout under `~/.vibe/worktrees/` and exit codes report the end, but there is no orchestrator and programmatic runs leave worktrees in place. source
Muse Code C
Meta Platforms, Inc.
The lead can request an isolated Git worktree per subagent, but there is no top-level worktree flag and the compatibility flag does not force isolation. source
OpenAI Codex BB
OpenAI OpCo, LLC
`codex exec --worktree` and `codex fork --worktree` give each run a managed worktree, but exit codes are undocumented and acc-08 hit tool-call limits in cloud containers within 20 minutes. source
OpenCode BB
Anomaly Innovations, Inc.
A `worktree.directory` setting and server sessions support parallel checkouts, but no run-level worktree flag or status report was found in the CLI pages. source
OpenMausBot B
Supamaus Software Private Limited
Bots run on their own computers and `run --json` reports which finished or stalled, but no git worktree or branch isolation is documented. source
Orca C
Lovecast Inc.
The README claims each agent runs in its own worktree with an orchestrator splitting tasks, but it is the vendor's claim without a documented flag or status report. source
Mirror CLI C
Reflection AI
`/worktree <name>` gives each agent a `mirror/<name>` branch under `~/.mirror/worktrees/` for parallel tabs, but orchestration is manual in the TUI and `exec` has no worktree flag. source
Rovo Dev CLI E
Atlassian Pty Ltd
`--worktree <name>` isolates each run under `.rovodev/worktrees/`, but an orchestrator learns the outcome only through server mode. source
Unstoppable E
Unstoppable Domains, Inc.
Parallel worktrees with review and Create PR are described in the support index, but only index lines and the changelog were read, and chats start in Full access. source
Verdent E
Verdent AI, Inc.
Workspace isolation uses git worktrees per workspace, but the page says manual workspace creation is hidden in the current desktop interface. source
Warp (Warp Agent CLI and the Oz Agent API) C
Denver Technologies, Inc. (d/b/a Warp)
Each cloud run gets its own environment and Oz can run Claude Code or Codex, but there is no worktree model for a local checkout. source

Couldn't check 3

The docs were closed to us, or silent. These carry no mark either way.

ListingWhy
Parallax E
Conifer Solutions, Inc.
The app runs agents side by side on the owner's subscriptions, but worktree isolation and status reporting are not in the pages read.
Qoder CLI C
BRIGHT ZENITH PRIVATE LIMITED
docs.qoder.com answered 502 from the fetch proxy on three attempts and a DNS failure through WebFetch, so no page was read.
Qwen Code BB
Qwen team, Alibaba Group (no legal entity is named in the repository or the docs)
Worktrees and Managed Session worktrees pages exist in the navigation but were not read.

Doesn't do this job 20

Listings in the category that do something else, with the reason.

ListingWhy
Agent Zero D
Agent Zero, s.r.o.
No worktree or multi-harness support; Agent Zero is one agent framework with sub-agents. source
Aider D
Aider AI LLC
No worktree flag; parallel runs share the checkout. source
Antigravity CLI C
Google LLC
No worktree flag or parallel-session command appears in the headless flag reference or permissions docs; each agy -p runs in the current directory. source
Auggie CLI D
Augment Computing, Inc.
No worktree flag in the CLI reference; parallel runs would share whatever checkout they are started in. source
AXLR BB
Underpass AI
No worktree support is documented; isolation comes from per-execution sandbox profiles, not separate checkouts. source
Crush D
Charmbracelet, Inc.
No worktree flag; the optional server mode shares workspaces rather than isolating checkouts. source
Pi B
Earendil Inc.
No worktree flag; parallel runs share the checkout they start in. source
goose BB
Agentic AI Foundation (Linux Foundation)
No worktree flag; sub-recipes run in the same directory. source
Harness² E
Google LLC
--workers runs benchmark domains concurrently; no worktrees or hand-offs between harnesses on a repository. source
Jules D
Google LLC
jules remote new --parallel started sessions in separate VMs, but new sessions stopped after 11 November 2026. source
Junie CLI D
JetBrains s.r.o.
No worktree flag in the CLI reference. source
Kimi Code CLI C
Moonshot AI (the global Kimi terms name NOVASCENT PRIVATE LIMITED, Singapore)
No `--worktree` flag or worktree docs; remote environments and subagents exist but share the session's checkout. source
Kiro CLI B
Amazon Web Services, Inc.
No worktree flag in the command reference; runs share the current checkout. source
Letta B
Letta
No worktree or branch isolation is documented; subagents run inside the same agent. source
mini-swe-agent B
No legal entity named. The MIT licence names Kilian A. Lieret and Carlos E. Jimenez
No worktree support; the local environment runs in the current directory. source
MiniMax Code B
Nanonoble Pte. Ltd.
No worktree flag among the `exec` options; `--cwd` picks an existing directory. source
OpenAI Agents API BB
OpenAI OpCo, LLC
Sessions get isolated sandboxes, not branches of one checkout, and no hand-off to other harnesses exists. source
OpenHands BB
All Hands AI
Conversations get their own Docker sandbox, not worktrees of one checkout, and no hand-off between harnesses is documented. source
Prime Agent C
Prime Intellect, Inc.
`rlm.spawn` subagents run in parallel inside one session; no worktree isolation is documented. source
Vibe-Trading C
HKUDS
No worktrees or coding agents. source

What stops agents finishing

  • orchestrator runs every child with permissions bypassed and no sandbox
  • tool-calling limits in cloud containers within 20 minutes (acc-08)
  • child harness shows a dialog the orchestrator cannot answer (acc-15)

Around it

Before: issue-tracking.list-issues. After: code-hosting.open-pull-request.

Where this comes from

Weight 4 of 100 in Agent harnesses: 2 of 33 accounts (acc-08, acc-15); 15 of 56 listings are orchestrators, so the weight is low on evidence and should rise when accounts appear

The same as data: /api/v1/use-cases/agent-harnesses.parallel-worktrees.json, or the MCP tool get_use_case with id agent-harnesses.parallel-worktrees. Other jobs in this category: /api/v1/niches?category=agent-harnesses.

Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person. A mark says what the docs show, not what a run measured.

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.