Niches · Agent harnesses

Fix a failing test or CI run and push the fix

No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.

0companies, full fit
40companies, fit with a limit
101companies carry the tag
18share of the category's work
5first-hand accounts

What counts as done

Changes are on a branch or commit the person can review, made in one invocation under a stated turn and time cap, and nothing else in the repository changed.

Measured, for our own tests: tests or CI pass after the run. No docs we read state these figures, so they don't count against a listing here.

How agents ask for it: CI is failing on main, fix it, make the tests pass, fix the lint errors and push, a PR has failing checks, repair it

Does the job, with a limit 42

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
Aider D
Aider AI LLC
aider --message with --yes-always, --test-cmd and auto-commits does one change and exits, but there is no MCP, no JSON output, .aider.conf.yml in the repository runs commands without a prompt, and no release since February 2026. source
Amp C
Amp Frontier Corporation
amp -x runs to completion and --stream-json ends with a result carrying is_error, and --mcp-config attaches a server per run, but Amp never asks before running tools, has no sandbox, and documents no exit codes. source
Antigravity CLI C
Google LLC
agy -p with action(target) rules under permissions.allow and a sandbox on by default covers the run, but a tool refused for lack of approval is soft-denied: the run continues, exits 0 and only prints to stderr. source
Auggie CLI D
Augment Computing, Inc.
auggie --print with --permission rules, --mode json and --mcp-config fits CI, but ask-user rules are denied without a UI, unmatched tools follow undefined implicit behaviour, there is no sandbox, and print mode may be off under an enterprise agreement. source
AXLR BB
Underpass AI
axlr-tui --exec --json with --policy, --budget and --sandbox files runs one prompt and prints a result line, but the service needs KMP and MADE engines, nothing has been run by us, and no GitHub or CI integration is documented. source
Claude Code C
Anthropic, PBC
claude -p takes --permission-mode, --allowedTools, --mcp-config and --max-turns and exits non-zero on failure, but a denied Write still gives exit 0 and is_error false (acc-12), and a read-only token stopped comments silently (acc-02). source
Clawboo B
Sanreds
The task runner works in a git worktree, runs the project's test command and a review agent, and the API can start runs, but Claude Code, Codex and Hermes teammates run with permissions off and no sandbox, and nothing stops edits elsewhere. source
CLI Manager BB
solhun (Jeong Kyeonghun)
The API opens a Claude Code or Codex session with a prompt, waits for idle and reads the screen, but the macOS app must be running, a permission prompt returns 409 awaiting_input for a person, and the result is screen text. source
Cline C
Cline Bot Inc.
cline "prompt" runs yolo by default with --json and --worktree, but the CLI has no allow or deny list for shell commands, only a PreToolUse hook, there is no sandbox, and exit codes are undocumented. source
Crush D
Charmbracelet, Inc.
crush run executes one prompt and exits, with MCP and permission deny rules from config, but it approves every permission request in the run, has no JSON output, no sandbox, and .crushrc in the repository executes at startup. source
Cursor CLI E
Anysphere, Inc.
agent -p with CURSOR_API_KEY, --trust and permissions in .cursor/cli.json runs in CI, but actions needing approval are rejected silently without --force, which then runs everything bar deny rules; print mode hung on macOS in early 2026 (acc-03). source
DeepSeek Harness (dsh) D
DeepSeek (the LICENSE names DeepSeek, and no legal entity is named in the repository or the docs)
dsh --profile headless --json --permission-mode workspace-write --add-dir <git dir> runs one task and exits 1 unless the final turn completed, but the session log goes to DeepSeek by default, reads and network are unconfined, and it is a developer preview. source
dsh-superagent (DeepSeek) E
DeepSeek (the LICENSE names the DeepSeek Harness Team; DeepSeek Harness's terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)
The default branch's CI is a trigger, each Ticket gets an Executor in its own git worktree and ends in a pull request, but Executors run without approval as the operator's OS user and it is a research prototype. source
Devin C
Cognition AI, Inc.
A session created over REST or MCP works in Devin's VM with a max_acu_limit and opens pull requests, but session creation has no idempotency key, a person must provision the service user, and usage has no published unit rate. source
Droid CLI C
The San Francisco AI Factory, Inc.
droid exec --auto medium runs installs, tests and commits, --auto high pushes, -o json gives is_error, and an action above the level exits non-zero, but a paid Factory plan and an fk- key from the web app come first. source
Pi B
Earendil Inc.
pi --print runs once and exits non-zero on an error stop reason, --tools limits tools and --no-approve ignores project config, but Pi has no permission system or sandbox and runs every tool call with the user's rights. source
Gemini CLI B
Google LLC
gemini -p with --approval-mode, --allowed-tools, GEMINI_TRUST_WORKSPACE and exit codes 0, 1, 42 and 53 fits CI, but a paid key or Code Assist licence is needed and the sandbox is off by default with network allowed on the default macOS profile. source
GitHub Copilot CLI C
GitHub, Inc.
copilot -p with --allow-tool, --deny-tool, --assisted-approval, --additional-mcp-config and --worktree runs headless on a GITHUB_TOKEN, but local sandboxing is off by default and exit codes are not tabulated. source
goose BB
Agentic AI Foundation (Linux Foundation)
goose run with --max-turns, --with-extension for a per-run MCP server and --output-format json fits CI, but the default mode approves every tool call and there is no sandbox. source
Junie CLI D
JetBrains s.r.o.
A one-shot run with JUNIE_API_KEY, --output-format json, --max-steps and an action allowlist fits CI, but the headless page is silent on what happens to an action needing approval, exit codes are undocumented, and project MCP servers load without asking. 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.)
kilo run --auto approves everything not denied, without it any rejected request exits 1, permission rules live in kilo.jsonc and exit codes are 0, 1 and 124, but the sandbox is off by default, MCP tools stop when network is restricted, and nothing caps turns or cost. source
Kimi Code CLI C
Moonshot AI (the global Kimi terms name NOVASCENT PRIVATE LIMITED, Singapore)
`kimi -p` runs without the TUI but always under Never Ask, so every write and shell command is auto-approved; deny rules in `[[permission.rules]]` are the only brake, there is no sandbox and no result status or cost in the output. source
Kiro CLI B
Amazon Web Services, Inc.
`kiro-cli chat --no-interactive --trust-tools=<ids>` with `permissions.yaml`, exit codes and stream-json cover the run, but the API key needs a paid plan, the sandbox is off unless `sandbox.json` enables it, and 2.0.x shipped consent prompts headless callers could not answer. source
Letta B
Letta
`letta -p --ephemeral --output-format json` runs headless with `--allowedTools` and `--disallowedTools`, but the default mode is `unrestricted`, there is no sandbox, no exit code table, and Letta Cloud is the default backend. source
mini-swe-agent B
No legal entity named. The MIT licence names Kilian A. Lieret and Carlos E. Jimenez
`mini -y --exit-immediately` with `cost_limit` and `step_limit` runs unattended, but the default local environment executes on the host with no isolation, there is no MCP client, and only the trajectory file records how the run ended. source
MiniMax Code B
Nanonoble Pte. Ltd.
`mcode exec --cwd --timeout --max-steps --output-format json` with exit codes does the run, but the default `smart` policy sends actions to a classifier and fails the run when a person would be needed, and no sandbox is documented. source
Mistral Vibe CLI B
Mistral AI (RCS Paris 952 418 325)
`vibe -p` with `--max-turns`, `--time-limit`, `--worktree` and exit codes 0 to 3 fits, but programmatic mode denies every approval-needing call unless `--auto-approve` or `permission = "always"` rules are set, and there is no sandbox on the host. source
Muse Code C
Meta Platforms, Inc.
`muse exec` with `META_API_KEY`, `--disable-approval`, `--max-model-steps` and the sandbox on fits, but exit 0 means the turn completed, not that tests pass, Linux runners need bubblewrap, and there is no free plan. source
OpenAI Agents API BB
OpenAI OpCo, LLC
A hosted sandbox with a network allowlist and MCP servers can run the task and turns report `failed` or `idle`, but no repository, token or pull request integration is documented and retries have no idempotency key. source
OpenAI Codex BB
OpenAI OpCo, LLC
`codex exec --sandbox workspace-write --json` runs in an OS sandbox with network off, but that default blocks `npm install` (acc-04) and acc-05 shows `network_access = true` ignored on macOS; exit codes are not documented. source
OpenCode BB
Anomaly Innovations, Inc.
`opencode run --standalone --auto --format json` runs in CI, but the base policy allows every action, there is no sandbox, exit codes are undocumented, and acc-09 shows a patch release breaking non-interactive runs. source
OpenHands BB
All Hands AI
The SDK or Agent Server runs a conversation headless with a confirmation policy and cost metrics, but the maintained paths are a library and a server, the CLI is retired, and git tokens come from the Cloud login. source
OpenMausBot B
Supamaus Software Private Limited
`openmausbot run TARGET TEXT --json --timeout` hands a job to a Claude Code or Codex bot and exits 0, 1 or 3, but the desktop app or server must be running, new bots start at Full access, and there is no sandbox. source
Paperclip C
Paperclip Labs, Inc.
A task wakes a Claude Code or Codex agent with `maxTurnsPerRun`, `timeoutSec` and an isolated worktree, but `claude_local` sets `dangerouslySkipPermissions` true by default and agents run on the host unless a sandbox provider is configured. source
Paseo B
Mohamed Boudra
`paseo run --provider codex --new-workspace worktree --output-schema` starts a Claude Code or Codex agent and waits, but permissions belong to the provider, the daemon must be running, and the run command's exit codes are not documented. source
Prime Agent C
Prime Intellect, Inc.
`-p --mode json` with `--autonomous-max-turns`, `--autonomous-max-tokens` and `--autonomous-timeout-ms` runs unattended, but there are no approvals and no sandbox, and the README says to use a disposable clone. source
Qwen Code BB
Qwen team, Alibaba Group (no legal entity is named in the repository or the docs)
`qwen -p` with `--approval-mode`, `--max-session-turns`, `--max-wall-time`, `--max-tool-calls` and exit 53 or 55 covers the run, but no sandbox is on by default, `--yolo` prints a warning rather than confining, and every run needs a paid key. source
Mirror CLI C
Reflection AI
`mirror exec` runs one session and `--max-iterations` caps it, but the directory's usage and folder notices must be accepted interactively first, calls needing approval fail unless `--approval-mode auto` is saved globally, and only a plain-text answer comes back. source
Rovo Dev CLI E
Atlassian Pty Ltd
`acli rovodev run [instruction]` runs once with `toolPermissions` allow, ask or deny and `--worktree`, but no JSON output, exit codes or turn cap are documented and the CLI needs a paid Rovo Dev plan an admin has activated. source
T3 Code C
T3 Tools Inc.
An outside agent can `t3_thread_launch` a Claude or Codex thread over MCP and `t3_thread_wait` on it, but new threads default to Full access, there are no caps, and the T3 server must be running. source
Traycer C
Traycer AI
The CLI creates worktrees and automates agents with `--json` NDJSON, but agents created from the CLI start in Full access unless `--permission-mode` is passed, and no caps or exit codes for agent runs were read. source
Warp (Warp Agent CLI and the Oz Agent API) C
Denver Technologies, Inc. (d/b/a Warp)
The `warp` CLI is interactive only; `POST /api/v1/agent/runs` with `WARP_API_KEY` runs a cloud or self-hosted agent whose `state` goes QUEUED, INPROGRESS, SUCCEEDED or FAILED, but it needs a Warp account and credits and the CLI has no sandbox. source

Couldn't check 2

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

ListingWhy
Orca C
Lovecast Inc.
The README says agents can start workers and create worktrees from the Orca CLI, but orca.dev disallows AI crawlers in robots.txt so the CLI reference was not read; no flags or exit codes seen.
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.

Doesn't do this job 12

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

ListingWhy
Agent Zero D
Agent Zero, s.r.o.
A Docker-hosted web agent with no print mode; the external API sends chat messages, not a one-shot run with a status, and there is no approval step or CI integration. source
Conductor E
Conductor
A Mac app with no headless mode; the HTTP API drives only cloud workspaces, which go read-only on 13 November and shut down on 20 November 2026. source
Emdash C
General Action, Inc.
A desktop app with no headless or print mode; the agents it runs do the work. source
Harness B
Autonomous, Inc.
Harness starts Claude Code or Codex in a tmux pane for a person to watch; it needs a browser sign-in and an opt-in setting, and is not itself a headless harness. source
Harness² E
Google LLC
A research CLI that runs OpenCode or Codex on JobBench, LAB and WorkBuddy-Bench tasks; no way to run it on your own repository is documented. source
HarnessRouter D
Lumentree Corporation
Tasks run in a hosted sandbox with files, artifacts and read-only environments; no repository clone, branch or push path appears in the run, harnesses or environments docs. source
Jules D
Google LLC
Sessions turned read-only after 11 November 2026 and the service ends on 31 December 2026, per the changelog, so no new work can finish. source
Parallax E
Conifer Solutions, Inc.
Parallax reviews finished sessions and opens fix PRs for proven findings; it does not take a failing CI run as a task. source
Sculptor C
Imbue, Inc.
A desktop app with no CLI, print mode or API. source
Unstoppable E
Unstoppable Domains, Inc.
No documented CLI, API or print mode; a Linux command-line tool appears only in changelog lines. source
Verdent E
Verdent AI, Inc.
No CLI, print mode or API; tasks start in the desktop app or a paired chat bot. source
Vibe-Trading C
HKUDS
A finance research agent; it does not take a repository or CI run as a task. source

What stops agents finishing

  • sandbox blocks package installs so tests cannot run (acc-04, acc-05)
  • a run reports success while the edit was denied (acc-12)
  • token scope too narrow to push or comment, found only after the run (acc-02)
  • paid plan or browser sign-in required before a headless key exists

Around it

Before: code-hosting.read-ci-logs, code-hosting.read-pr-diff. After: code-hosting.open-pull-request, code-hosting.comment-on-pr.

Where this comes from

Weight 18 of 100 in Agent harnesses: 6 of 33 accounts (acc-01, acc-02, acc-07, acc-08, acc-12, acc-27) and the first example in three vendors' headless docs

The same as data: /api/v1/use-cases/agent-harnesses.fix-failing-ci.json, or the MCP tool get_use_case with id agent-harnesses.fix-failing-ci. 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.