Learn · Agentic architecture & orchestration · CCAR-F task 1.4

Enforcement and handoff

Prompts are probabilistic, code is deterministic. Put must-never-break rules in code, and hand off with a summary that stands alone.

Key ideas

Prompts are probabilistic, code is deterministic: enforce must-hold rules in code.
A prerequisite gate inside the tool refuses to act until an earlier step succeeded.
Return a structured error that tells the model what to do next.
A state machine tracks the workflow stage and allows only valid transitions.
Split multi-concern requests, investigate each (in parallel if independent), answer together.
A human handoff summary must stand alone: customer ID, root cause, amounts, actions taken, recommendation.

The rule lives in code the model can't talk its way around. The prompt still says "verify first", so the model usually does; the code guarantees it.

Fixing a must-never-happen failure by strengthening the prompt ("ALWAYS verify first") instead of adding a code gate.
Returning a bare error with no hint, so the model can't recover.
Handing off to a human with "see the conversation above": the human can't see it.
Answering only the first concern in a multi-concern message.