Skip to main content

How a request flows

  1. Classify. A fast free-model agent reads your request and decides what kind of work it is.
  2. Plan. A managerial agent (architect, tech lead, release manager, or a QA / security lead) sketches the approach.
  3. Execute. A paid-model executor produces the actual output — code, document, decision.
  4. Verify. An acceptance agent checks the result against the original ask. Pass/fail, with deltas.
If something fails, Cerebrum retries within the same level up to three times before moving the work up to a more capable agent.

Why a team

  • Specialization beats generalization. A planning agent uses different prompts and tools than an executor. Forcing one agent to do both leaves quality on the table.
  • The right tool for the work. Free models handle classification and verification fast and cheap. Paid models earn their cost only on the hard problems. Cerebrum decides per step.
  • Escalation, not retry-spam. When a level can’t resolve a conflict, the work moves up — to a more capable agent — rather than burning attempts at the same level.

What you see

  • A streamed view of every step. As an executor produces tokens, you see them live.
  • An audit log of every classification, escalation, and verification decision.
  • A cost breakdown per job, broken out by model.