Run the Agent Review Loop

Connect an agent to one workspace review policy and request review from the people your workspace invites.

Agents start with Rating AI review when the owner connects them; invite reviewers when people should decide. Its ratings are advisory, separate from human-review requests, and cannot satisfy a required human review. Supported AI tasks.

Connected workspace review

The workspace MCP at https://www.rateloop.ai/api/agent/v1/mcp binds an approved agent version to its owner-approved review policy. A safe connection can read that context and evaluate review requirements, but it cannot read private artifacts or administer the workspace.

1. Call rateloop_get_agent_context, then rateloop_verify_connection.
2. Before each eligible output, call rateloop_evaluate_review_requirement with privacy-safe execution metadata.
3. If review is required, call rateloop_request_review, then rateloop_wait_for_review and rateloop_get_review_result before the host releases the output.
4. Call rateloop_get_assurance_state when you need current coverage and agreement evidence.
5. Call rateloop_get_model_guidance before work starts to see which model met the owner's review standard for each workflow. It is decision support from human reviews, never a review.

When Rating AI is enabled

This path supports custom binary evaluations of English or German text with the owner's saved question and answer labels. Explicit AI processing permission and a compatible evaluator are required. For AI-primary workflows, training is optional. Do not override the saved question or labels.

Full human review can also use optional blind AI comparisons when private learning is enabled in Improve. AI results remain withheld until human review finishes. These comparisons use the existing AI allowance and never add human assignments.

If selection.ratingAi.enabled is true, send the exact text to the authorized evaluator. Choose optional context or reference material and the required text to evaluate. Use an empty context string when none is needed. Hash those exact UTF-8 strings into sourceEvidence.hash and suggestionCommitment, then send the same strings in ratingInput when calling rateloop_evaluate_review_requirement:

ratingInput: {
  language: "en", // or "de"; match the owner's saved question language
  sourcePayload: contextOrReference ?? "",
  suggestionPayload: textToEvaluate,
}

Omit ratingInput when selection.ratingAi.enabled is not true. For AI-primary workflows, missing permission, an unavailable compatible model, missing text, or mismatched commitments produces an error; do not treat it as a completed rating. At full human review, unavailable AI or absent comparison input skips the optional comparison while human review continues. A policy result of review_not_required can still include a queued ratingAi job. Use its opportunityId with rateloop_get_rating_ai_result.

• AI only: poll queued or evaluating work with bounded waits. Report the result to the owner as advice, never as a human approval; do not create a human fallback at the AI-only endpoint.
• AI + human: complete rateloop_request_review, rateloop_wait_for_review, and rateloop_get_review_result. A selected blind AI comparison stays hidden until the human outcome. AI cannot satisfy required human review.
• Human only: no AI job is created. Follow the human review decision and result path.
• Offline, failed, cancelled, or uncertain: queued work may wait for the authorized local worker. Recheck the same job; report a terminal failure or uncertainty as no AI approval and let the accountable owner decide. Never turn an unavailable AI result into a human approval.

Payload limits

Human review accepts up to 131,072 characters per field. Each complete serialized MCP request must also fit 65,536 UTF-8 bytes (64 KiB), including both fields, JSON escaping and the envelope. Check the evaluation and eventual review request sizes before creating an opportunity. Reuse the exact committed text; never attach a summary's rating to a longer answer.

Rating AI supports at most 3,000 characters per field and the model's token limit. Longer material within the human material and transport limits returns AI status unavailable with rating_input_too_long and no AI job. Human review continues under the owner's policy on the same opportunity and card. Do not skip it because AI cannot rate the length, create an extra review, or add a human fallback to an AI-only policy.

MCP integrations are advisory: the agent must call the review tools, and your workflow controls when the output is delivered.

Connect with Codex

Open Connections, choose Codex desktop and copy its connection message. It names RateLoop Workspace so Codex can use it when the plugin is already available in the host's native inventory. RateLoop does not currently publish a public plugin marketplace or repository install source. If Codex offers no native plugin action, report the host blocker instead of attempting a repository install.

The copied message targets rateloop-workspace@rateloop. If Codex installs that already-available inventory entry, its skills and tools become available in one new Codex task. An already-loaded plugin can continue in the same task after OAuth.

Plugin installed, but the tools are missing?

• If the plugin is installed during the current task, start exactly one new Codex task and paste the same connection message once while its URL is still unclaimed.
• If an already-loaded plugin completes native OAuth before a claim, continue in the same task and do not paste the message again.
• After invalid_grant, stop RateLoop calls, never reuse the old connection URL, recover only through host-presented plugin OAuth, and then paste one new targeted reconnect message into one fresh task.
• Complete reconnect source confirmation and owner approval in the task that holds the fresh reconnect message; do not paste it again.
• After verification, future tasks rediscover the connection with rateloop_get_agent_context and do not need the connection message.
• One active RateLoop OAuth grant renews short-lived access across tasks for up to 90 days; an older task must never revoke that grant.

A verified OAuth grant works across tasks and renews short-lived access automatically while it remains active. Fresh authorization is required when the grant expires or is revoked, the workspace is deleted, or Codex restores a stale credential. Follow only an OAuth action presented by the plugin host. If Codex shows none, report the host blocker; standalone MCP login or logout does not repair plugin authentication.

A standalone MCP registration named rateloop-workspace must not coexist with the RateLoop Workspace plugin. If a new post-install task still lacks the protected tools, inspect the native plugin inventory once; do not reinstall an enabled plugin.

When a call is refused

Every failure returns the same envelope, and whether an agent should retry is stated in the response rather than inferred from the status code. API errors lists each code, what produced it, and which ones are retryable.

Decision boundary

RateLoop supplies human-review evidence and disclosed limitations; it does not issue an automatic production, safety, legal, medical, or compliance approval. Use only authorized, minimized material and keep an accountable person responsible for every rollout decision.