Clarovo

Responsible AI

What our sub-agents do, what they don't, the harness that wraps them, and the methodology coverage metric you can defend.

What sub-agents do

What sub-agents do

14 orchestrator-driven sub-agents run inside the Harness. Each does one thing, in one order, with one input and one output.

The sub-agents are: situation_assessor, answer_interrogation, consequence_engine, avoidance_question, bottleneck_diagnostician, market_signal_analyzer, pre_mortem_runner, stakeholder_simulator, competitor_watcher, benchmark_puller, scenario_engineer, burn_rate_calculator, decision_memo_writer, clarity_mapper. Each is a LangGraph StateGraph with a Pydantic state model. Nodes are either deterministic Python or structured-output LLM calls. No free-form chat. No agent has free-form intelligence.

What sub-agents don't do

What sub-agents don't do

  • They don't proceed without a sign-off at the supervised gates. The engagement pauses for the EM to review + approve / edit / flag at 5-10 of the 50+ sub-agent invocations.
  • They don't return half-formed artifacts. If the LLM call returns malformed output, the sub-agent retries up to 3 times; if it still fails, the run is marked failed and the EM sees a "the system couldn't produce a confident answer" message.
  • They don't exceed the token budget. Every sub-agent has a per-call budget. If exceeded, the sub-agent stops and surfaces a "this analysis is too expensive, give me a smaller scope" message.
  • They don't make claims that aren't grounded in the source material. Every answer is post-processed through the citation chain: fact #N in engagement_facts → Q of the session → answer (confidence 0.X) → Regret Score N → memo paragraph.
The harness

The 7 principles every sub-agent must satisfy

  1. Deterministic agents. One input, one output, one order. No free-form chat.
  2. Input validation gate. Bad input → fail fast with a clear error. The sub-agent never starts with malformed data.
  3. Output validation gate. Bad output → retry up to 3 times, then mark the run failed.
  4. Supervised gates. The EM signs off at 5-10 points across the engagement. The sub-agent does not proceed without sign-off.
  5. Audit trail. Every fact cited, every answer produced, every confidence score, every sign-off decision is timestamped + attributed + traceable to the ingested source material.
  6. Rollback. If the EM rejects a sign-off, the harness rolls back the sub-agent's writes to the canvas. Reversible until signed off.
  7. Cost + token budget. Every sub-agent has a budget. Exceeded → stop + surface the message.
The methodology coverage metric

The methodology coverage metric (the partner-attestation artifact)

The Harness tracks which methodology steps ran, which didn't, and the overall coverage percentage per engagement. The MBB rigor stamp (the RigorStamp component on the orchestrator panel) reads from this. The partner can say "this engagement has 92% methodology coverage" and the client can verify it. Without the harness, this number is fake. With the harness, it's the literal count of methodology steps executed vs. required.

Your data

Your data

Supabase Postgres + RLS. Every table has a workspace-scoped RLS policy. No cross-workspace reads, ever. The LLM provider is OpenRouter; we don't train on your data. The embedding store is pgvector; embeddings are workspace-scoped. We delete a workspace's data within 30 days of the delete request.

The full privacy surface is at /privacy. Report a concern to responsible-ai@clarovo.xyz.