# Drive the harness from another program and gate on its result

> Drive the harness from another program and gate on its result: 8 companies fully do it, 35 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.embed-in-program/
- Markdown: https://www.anchorterminal.com/niches/agent-harnesses.embed-in-program/index.md (~3,800 tokens)
- Slim: https://www.anchorterminal.com/niches/agent-harnesses.embed-in-program/index.min.md (~830 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/agent-harnesses.embed-in-program/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

Four or more companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 8 | 35 | 77 | 14 | 8 |

## What counts as done

Each run exits on its own with a machine-readable result that says whether the task was done and what was denied, and a parent process can tell a failed run from a finished one without reading prose.

How agents ask for it: run the agent on each item in this queue; call the coding agent from my orchestrator; start a session over HTTP and read the result; let my agent hand a task to Codex

## Does the job

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

- **AXLR (Underpass AI)**: --exec --json ends with a result line, exit 3 means a call waits for --decide approve or deny, forbidden calls are marked in tool_calls, and the HTTP, gRPC and MCP service takes Idempotency-Key per turn.
- **Droid CLI (The San Francisco AI Factory, Inc.)**: -o json returns is_error, num_turns, session_id and usage, an action above the autonomy level stops the run with a stderr error and a non-zero exit, exit codes 0, 1 and 2 are documented, and stream-jsonrpc drives multi-turn sessions.
- **MiniMax Code (Nanonoble Pte. Ltd.)**: `ExecResult` JSON carries `status` (succeeded, failed, timeout, cancelled, limit_exceeded), exit codes distinguish input, config, task, timeout and limit failures, and a pending permission request fails the run instead of hanging.
- **Mistral Vibe CLI (Mistral AI (RCS Paris 952 418 325))**: Exit 0 finished, 1 could not start, 2 model or runtime failure, 3 stopped by a limit or refusal; `--output json|streaming` and `export.json` give outcome, stop reason, tokens and cost.
- **OpenAI Agents API (OpenAI OpCo, LLC)**: Sessions and turns carry `status` and `error` with named codes, `agent.session.turn.failed` and `requires_action` tell a parent what happened, and SDKs plus webhooks wrap it.
- **OpenHands (All Hands AI)**: `conversation.run()` completes and exits, `set_confirmation_policy` and `reject_pending_actions` control approvals, and `conversation_stats` reports accumulated cost and tokens.
- **OpenMausBot (Supamaus Software Private Limited)**: `--json` prints one document with the job `state`, exit 3 means someone must answer, and the MCP server's `wait_for_conversation` returns settled, needs-user, failed, stalled or timed-out with a `send_id` for idempotent sends.
- **Paperclip (Paperclip Labs, Inc.)**: A REST API with `/api/openapi.json`, board and agent tokens, issue status fields, and a CLI whose `--json` prints the raw API payload.

## 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.

- **Agent Zero (Agent Zero, s.r.o.)**: POST /api_message with X-API-KEY sends a message and api_log_get reads the log, but the docs defer request and response shapes to the source and no done-or-denied field is described.
- **Amp (Amp Frontier Corporation)**: The final result message has subtype, is_error and num_turns, --stream-json-input drives several turns, and SDKs exist, but nothing is ever denied so there is no denial field, and exit codes are undocumented.
- **Antigravity CLI (Google LLC)**: json and stream-json end with a result whose status is SUCCESS, ERROR, CANCELED or WAITING and a failed run exits 1, but a denied tool leaves status SUCCESS and exit 0 with only a stderr notice.
- **Auggie CLI (Augment Computing, Inc.)**: --mode json emits structured events and a TypeScript SDK exists, but exit codes are undocumented and the docs do not say how a denied ask-user call appears in the output.
- **Claude Code (Anthropic, PBC)**: json and stream-json end with a result carrying is_error, permission_denials and total_cost_usd, and --permission-prompts none denies instead of waiting, but exit 0 and is_error false when the edit was denied (acc-12); .claude writes denied silently (acc-13).
- **Clawboo (Sanreds)**: A REST API behind an access token starts runs, answers approval cards and lists budgets, but no machine-readable done-or-denied result per run was found in the docs read.
- **CLI Manager (solhun (Jeong Kyeonghun))**: REST and MCP on 127.0.0.1 with a bearer token, HTTP statuses for awaiting_input, disconnected and timedOut, but the outcome is screen text and session state, not a done-or-denied field.
- **Cline (Cline Bot Inc.)**: --json emits messages as JSON and -P picks a provider, but exit codes are undocumented, nothing is denied in the default yolo mode, and CLI 3.0.70 removed options without notice.
- **Cursor CLI (Anysphere, Inc.)**: json and stream-json end with a result event, but a rejected action leaves no field in the output, exit codes are undocumented, and print mode hung with no output for over 90 seconds (acc-03).
- **DeepSeek Harness (dsh) (DeepSeek (the LICENSE names DeepSeek, and no legal entity is named in the repository or the docs))**: --json opens with a session event and closes with final, exit 0 means completed and 1 aborted or errored, and AUTO_REVIEW_DENIED marks a policy denial, but it is a developer preview whose README warns of breaking changes.
- **dsh-superagent (DeepSeek) (DeepSeek (the LICENSE names the DeepSeek Harness Team; DeepSeek Harness's terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.))**: A Management API with revisions and a status script exist, but tasks arrive through GitHub conversations rather than a program call, and the result is a pull request, not a per-run record.
- **Devin (Cognition AI, Inc.)**: Sessions have states, devin_session_gather waits for finished, errored, sleeping or waiting, and max_acu_limit bounds each, but creation is not idempotent so a retried timeout doubles the work.
- **Pi (Earendil Inc.)**: --mode json streams JSONL to agent_settled and RPC mode takes commands on stdin, but a failed response does not set a non-zero exit in JSON mode and nothing is ever denied.
- **Gemini CLI (Google LLC)**: Exit codes 0, 1, 42 and 53 and a JSON envelope with response, stats and error are documented, but how a tool denied in headless mode appears in the output is not; the TUI, not -p, hung on bare Ubuntu (acc-06).
- **GitHub Copilot CLI (GitHub, Inc.)**: -p with --output-format json ends in a result record with an exit status even on early failure, but exit codes are not documented as a table and patch releases have removed flags.
- **goose (Agentic AI Foundation (Linux Foundation))**: json and stream-json output and exit codes to check are documented, but no code table exists, a run that ends with a question has not completed, and MCP transports closed under Ollama (acc-18).
- **Harness (Autonomous, Inc.)**: harness new, wait and show take --json and wait exits 0 idle, 3 needs the person, 4 stopped, 124 timed out and 5 refused, but the person must sign in and turn on Agents can start other agents first, and the result is screen text.
- **HarnessRouter (Lumentree Corporation)**: POST /{harness_id}/v1/responses with an Idempotency-Key streams SSE events with response and session ids, cancel stops work and budgets are per turn, but closing the stream leaves the agent running and billing, and a person must top up credit first.
- **Junie CLI (JetBrains s.r.o.)**: json and json-stream output exist, but exit codes are undocumented and what a headless run does when an action needs approval is not stated.
- **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.))**: --format json streams events, exit 0, 1 and 124 are documented, and an auto-rejected request without --auto exits 1 with a stderr cause, but 7.8.13 and 7.8.18 changed permission and MCP behaviour in patch releases.
- **Kimi Code CLI (Moonshot AI (the global Kimi terms name NOVASCENT PRIVATE LIMITED, Singapore))**: `-p --output-format stream-json` emits one JSON object a line, plus ACP and a local REST server with OpenAPI, but no final result event with status or cost and no exit code table are documented, and denials cannot occur because print mode auto-approves.
- **Kiro CLI (Amazon Web Services, Inc.)**: stream-json events and exit codes 0, 1, 3 and 4 are documented, but a denied tool is reported as a `[denied]` text line, not a field, and acc-15 and acc-19 show flags that opened interactive prompts under an orchestrator until 2.0.1.
- **Letta (Letta)**: JSON result with `usage` and a bidirectional stream-json mode with `can_use_tool` control requests, but no documented exit codes or denial field for non-interactive runs.
- **mini-swe-agent (No legal entity named. The MIT licence names Kilian A. Lieret and Carlos E. Jimenez)**: It is a Python library with a `-o` trajectory file, so a parent can parse the trajectory, but process exit codes and a denial field are not documented.
- **Muse Code (Meta Platforms, Inc.)**: `--json` JSONL and exit codes 0, 1, 2, 130 and 143 exist, but a held approval in a headless run stays unresolved rather than failing, and the events carry no cost.
- **OpenAI Codex (OpenAI OpCo, LLC)**: `--json` emits `turn.completed` and `turn.failed` with usage, `--output-schema` fixes the final shape, and TypeScript and Python SDKs exist, but exit codes and the `error` payload are undocumented.
- **OpenCode (Anomaly Innovations, Inc.)**: `run --format json` gives NDJSON and the server has an OpenAPI spec, but the final event and exit codes are undocumented, and acc-09 and acc-10 are programs failing on the non-interactive and server paths.
- **Parallax (Conifer Solutions, Inc.)**: `GET $PARALLAX_URL/v1/sweeps` and `/v1/calibration` answer a bearer capture token with JSON, but the base URL is unpublished and there is no OpenAPI or endpoint reference.
- **Paseo (Mohamed Boudra)**: The `@getpaseo/client` SDK creates agents over WebSocket and the CLI has `--json` and `--output-schema`, but only the usage commands document exit statuses.
- **Prime Agent (Prime Intellect, Inc.)**: The json mode streams session events and `--mode rpc` and `acp` exist, but exit codes beyond 1 for a failed command and a denial field are not documented.
- **Qwen Code (Qwen team, Alibaba Group (no legal entity is named in the repository or the docs))**: JSON and stream-json output with usage, exit 53 for the turn cap, 55 for a budget, 52 for bad config and 130 for SIGINT, but no documented field says what was denied.
- **Mirror CLI (Reflection AI)**: `mirror exec` prints plain text with undocumented exit codes; `mirror acp` has a structured protocol, but both need the interactive notice step first.
- **Rovo Dev CLI (Atlassian Pty Ltd)**: `acli rovodev serve <port>` streams `/v3/stream_chat` as server-sent events and `pause_on_call_tools_start=true` lets a program approve or deny each tool call, but the one-shot `run` has no JSON or exit codes.
- **T3 Code (T3 Tools Inc.)**: The MCP server lets outside agents start, message, wait on and read threads with a sign-in approval flow, but no result status, denial field or exit code is documented.
- **Traycer (Traycer AI)**: `--json` writes NDJSON for every CLI command and `host capabilities --has` exits non-zero, but the agent automation commands were not read in detail.
- **Vibe-Trading (HKUDS)**: `vibe-trading run -p ... --json`, a REST API behind `API_AUTH_KEY` and a stdio MCP server exist, but no exit code table or denial field is documented.
- **Warp (Warp Agent CLI and the Oz Agent API) (Denver Technologies, Inc. (d/b/a Warp))**: The platform API returns `run_id` and `state`, error bodies carry `code` and `retryable`, and a TypeScript SDK exists, but what was denied is not reported and a Warp account is required.

## Couldn't check

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

- **Orca (Lovecast Inc.)**: An Orca CLI exists per the README; its output format and exit codes were 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: Aider (Aider AI LLC): The scripting page has --message and a Python API that print text; there is no JSON result, denial field or documented exit code, so a parent must read prose or git.
- not this job: Conductor (Conductor): The beta API at api.conductor.build/v0 creates cloud workspaces and polls session status, and the cloud goes read-only on 13 November 2026 and is deleted on 20 November.
- not this job: Crush (Charmbracelet, Inc.): crush run prints text to stdout with --quiet to hide the spinner; there is no JSON mode, no denial field and no documented exit code, so a parent cannot tell done from failed without reading prose.
- not this job: Emdash (General Action, Inc.): No API or CLI for another program; it is a GUI.
- not this job: Harness² (Google LLC): The CLI accepts only the three benchmarks; a program cannot hand it its own task.
- 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: Sculptor (Imbue, Inc.): No API, CLI or MCP server.
- not this job: Unstoppable (Unstoppable Domains, Inc.): No API, MCP server or CLI is documented for outside callers.
- not this job: Verdent (Verdent AI, Inc.): No API, SDK or CLI; the only outside entry is a paired Telegram, Slack or Feishu bot.

## What stops agents finishing

- print mode hangs with no output and no timeout (acc-03)
- a consent dialog appears on first run and cannot be answered (acc-15, acc-19)
- exit 0 and is_error false when the edit was denied (acc-12)
- patch release breaks the non-interactive path (acc-09)
- no batch mode for a headless server (acc-10)

## Around it

Before: agent-frameworks.plan-task, queues.dequeue. After: agent-harnesses.fix-failing-ci, code-hosting.open-pull-request.

## Where this comes from

Weight 14 of 100 in Agent harnesses: 8 of 33 accounts (acc-03, acc-06, acc-09, acc-10, acc-12, acc-13, acc-15, acc-19) are programs calling a harness and failing on its interface

As data: https://www.anchorterminal.com/api/v1/use-cases/agent-harnesses.embed-in-program.json, or the MCP tool get_use_case with id agent-harnesses.embed-in-program. 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
