# Build or port a whole feature from a spec in an unattended loop

> Build or port a whole feature from a spec in an unattended loop: 0 companies fully do it, 39 with a limit, 77 carry the tag. Every listing in Agent harnesses with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/agent-harnesses.spec-to-build-loop/
- Markdown: https://www.anchorterminal.com/niches/agent-harnesses.spec-to-build-loop/index.md (~2,800 tokens)
- Slim: https://www.anchorterminal.com/niches/agent-harnesses.spec-to-build-loop/index.min.md (~730 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/agent-harnesses.spec-to-build-loop/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

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

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 0 | 39 | 77 | 12 | 4 |

## What counts as done

Every story in the spec has a commit, the loop stopped on its own at the completion marker or iteration cap, and total spend stayed under the stated budget.

Measured, for our own tests: every story in the spec has a passing test.

How agents ask for it: work through prd.json until every story passes; port this Python tool to Go; build the app from the spec overnight; keep going until the tests pass

## Does the job, with a limit

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.

- **Aider (Aider AI LLC)**: A shell loop over --message works and each edit is a commit, but nothing caps turns or spend.
- **Amp (Amp Frontier Corporation)**: A loop can wrap amp -x and read is_error, but no turn, time or cost cap exists and usage is reported in tokens only; background subagents keep running after the process exits.
- **Antigravity CLI (Google LLC)**: A loop can drive agy -p and read status, but no turn cap or spend cap exists, only --print-timeout, and usage is tokens without cost.
- **Auggie CLI (Augment Computing, Inc.)**: Print mode and --show-cost give a scriptable loop with cost in logs, but no turn or spend cap exists in the 1.x reference.
- **AXLR (Underpass AI)**: Per-execution budgets in usd, seconds and tool_calls and max_session_usd bound spend, but no built-in loop or completion marker exists and the project is pre-1.0.
- **Claude Code (Anthropic, PBC)**: A bash loop over claude -p built two apps in a weekend (acc-28) and --max-turns and --max-budget-usd bound each run, but the budget counts estimated spend (acc-16) and the loop exhausted two Max 20x subscriptions.
- **Clawboo (Sanreds)**: Per-task worktrees, verification and hard-cap USD budgets fit a long build, but there is no completion-marker loop and the runtimes run unattended with permissions bypassed.
- **Cline (Cline Bot Inc.)**: --worktree and --retries help a loop, but no turn or spend cap and no cost field were found.
- **Crush (Charmbracelet, Inc.)**: A loop can wrap crush run, but nothing caps turns or spend and no cost figure comes back per run.
- **Cursor CLI (Anysphere, Inc.)**: A loop can use agent -p --force, but no turn or spend cap exists and the result carries duration, not cost.
- **DeepSeek Harness (dsh) (DeepSeek (the LICENSE names DeepSeek, and no legal entity is named in the repository or the docs))**: Resume with --session-id and a documented supervisor workflow exist, but no turn or spend cap and no cost field were found.
- **Devin (Cognition AI, Inc.)**: Playbooks and max_acu_limit per session bound a long task, but there is no loop with a completion marker and spend is in ACUs without a published rate.
- **Droid CLI (The San Francisco AI Factory, Inc.)**: --worktree and --fork help a loop, but no flag caps turns or spend, and rolling rate limits are unstated.
- **Pi (Earendil Inc.)**: Sessions and --mode json support a loop wrapper, but nothing caps turns or spend.
- **Gemini CLI (Google LLC)**: maxSessionTurns and exit 53 bound a loop, but no spend cap exists and stats are tokens only.
- **GitHub Copilot CLI (GitHub, Inc.)**: --autopilot keeps working until task_complete and /goal --max-ai-credits caps credits, but the credit cap is interactive only and -p has no turn cap.
- **goose (Agentic AI Foundation (Linux Foundation))**: GOOSE_MAX_TURNS and GOOSE_SUBAGENT_MAX_TURNS cap turns including subagents, but no spend cap or cost output exists.
- **HarnessRouter (Lumentree Corporation)**: Per-turn defaults of 400 steps, 7200 seconds and 10 USD bound each turn and sessions continue, but no loop or completion marker exists and disabled_tools is only an instruction on some bases.
- **Junie CLI (JetBrains s.r.o.)**: --max-steps bounds each task, but no spend cap or cost output exists.
- **Kilo Code CLI (Kilo Code Inc. (the terms are a contract with Anaconda, Inc. on behalf of itself and its affiliates, including Kilo Code Inc.))**: --worktree and sessions help, but no turn, time or cost cap exists for kilo run.
- **Kimi Code CLI (Moonshot AI (the global Kimi terms name NOVASCENT PRIVATE LIMITED, Singapore))**: A step cap exists per turn, but no dollar budget, no cost output and no worktree support; the loop itself needs an outside wrapper.
- **Kiro CLI (Amazon Web Services, Inc.)**: No turn cap, spend cap, cost output or worktree flag; a loop wrapper gets exit codes only.
- **Letta (Letta)**: No turn cap, spend cap or worktree; token usage is reported but no cost.
- **mini-swe-agent (No legal entity named. The MIT licence names Kilian A. Lieret and Carlos E. Jimenez)**: Cost and step limits exist, but no worktree, no built-in loop, and the limits are LiteLLM estimates.
- **MiniMax Code (Nanonoble Pte. Ltd.)**: Step and time caps exist, but no dollar budget, token usage only, and no worktree flag.
- **Mistral Vibe CLI (Mistral AI (RCS Paris 952 418 325))**: `--max-turns`, `--max-tokens`, `--time-limit` and `--worktree` are there, but the docs say not to rely on `--max-price` for hard budget enforcement.
- **Muse Code (Meta Platforms, Inc.)**: `--max-model-steps` caps work and isolated subagents get worktrees, but no dollar budget and no cost output.
- **OpenAI Agents API (OpenAI OpCo, LLC)**: OpenAI runs the loop with subagents and a session budget code, but hosted sandboxes expire after an hour idle and no per-turn cost is shown.
- **OpenAI Codex (OpenAI OpCo, LLC)**: `--worktree` and `exec resume` support a loop, but there is no turn or spend cap and acc-26 needed about six interventions and hit quota over a month of porting work.
- **OpenCode (Anomaly Innovations, Inc.)**: No turn or spend cap; worktree placement is configurable but there is no cost output to stay under a budget.
- **OpenHands (All Hands AI)**: Iterative refinement and a stuck detector are guides, and costs are tracked, but no per-run spend cap exists in the SDK pages read.
- **OpenMausBot (Supamaus Software Private Limited)**: `--timeout`, `--strict` and job trees bound a run, but there is no spend cap or cost output.
- **Paperclip (Paperclip Labs, Inc.)**: Heartbeats, `maxTurnsPerRun` and budget hard stops bound the work, but the docs say enforcement uses recorded spend so in-flight work can delay a stop.
- **Paseo (Mohamed Boudra)**: No turn or spend cap; usage limits are visible per account but not enforced.
- **Prime Agent (Prime Intellect, Inc.)**: `/autonomous` continues within turn, token and time budgets with quality gates, but the README says reaching a limit does not imply success and there is no spend cap in dollars.
- **Qwen Code (Qwen team, Alibaba Group (no legal entity is named in the repository or the docs))**: Three run budgets exist, but `--max-tool-calls` does not count a subagent's inner calls and there is no dollar budget or cost output.
- **Mirror CLI (Reflection AI)**: `--max-iterations` and worktrees exist, but no budget, cost output or built-in loop.
- **Rovo Dev CLI (Atlassian Pty Ltd)**: Worktrees and permissions exist, but no turn or spend cap and no cost output.
- **Traycer (Traycer AI)**: Usage limits are read from providers and routed around, but nothing caps turns or spend.
- **Verdent (Verdent AI, Inc.)**: Manager plans and dispatches Worker tasks, including to Claude Code or Codex through BYOA, but no caps, cost output or completion marker are documented.
- **Warp (Warp Agent CLI and the Oz Agent API) (Denver Technologies, Inc. (d/b/a Warp))**: Cloud runs and budgets bound spend at the platform level, but no turn cap or per-run cost was found.

## Couldn't check

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

- **Orca (Lovecast Inc.)**: No caps or cost reporting seen; docs not read.
- **Qoder CLI (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.

- not this job: Agent Zero (Agent Zero, s.r.o.): Not aimed at unattended repository builds; tasks and goals are chat-scoped.
- not this job: CLI Manager (solhun (Jeong Kyeonghun)): No loop, budget or completion handling; it starts and watches sessions.
- not this job: Conductor (Conductor): No loop or budget functions; the app orchestrates sessions a person reviews.
- not this job: dsh-superagent (DeepSeek) (DeepSeek (the LICENSE names the DeepSeek Harness Team; DeepSeek Harness's terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)): One Ticket per request with a 30-minute background limit; not built to work through a spec until done.
- not this job: Emdash (General Action, Inc.): No loop, budget or completion handling.
- not this job: Harness (Autonomous, Inc.): No loop, budget or completion handling.
- not this job: Harness² (Google LLC): Refines a harness over benchmark tasks with --k steps; not a spec-to-build loop.
- not this job: Jules (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.
- not this job: Parallax (Conifer Solutions, Inc.): Not a build loop; it reviews sessions other harnesses ran.
- not this job: Sculptor (Imbue, Inc.): Interactive only; no caps or unattended loop.
- not this job: T3 Code (T3 Tools Inc.): No turn or spend caps and no built-in loop.
- not this job: Unstoppable (Unstoppable Domains, Inc.): No caps or unattended loop documented.
- not this job: Vibe-Trading (HKUDS): Not a build loop.

## What stops agents finishing

- two Max 20x subscriptions exhausted in days (acc-28)
- a day of Hermes cost $300 (acc-08)
- no clear endpoint so the loop stops early or never (acc-28)
- critic sub-agent misses invalid output the sandbox cannot render (acc-07)

## Around it

Before: planning.write-spec. After: agent-harnesses.fix-failing-ci, code-hosting.open-pull-request.

## Where this comes from

Weight 12 of 100 in Agent harnesses: 4 of 33 accounts (acc-07, acc-08, acc-26, acc-28); the Ralph loop is the most written-about pattern of early 2026 but most posts are guides, not runs

As data: https://www.anchorterminal.com/api/v1/use-cases/agent-harnesses.spec-to-build-loop.json, or the MCP tool get_use_case with id agent-harnesses.spec-to-build-loop. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
