@ichiba
3 layers, every unbiased.
Onerous constraints according to agent. Each and every agent has an specific “What You By no means Do” segment — no longer tips, exact laws. The Complete Stack agent can not invent an auth type the Architect did not outline. The DevOps agent can not skip checks. The Safety agent has 0 code-writing gear — learn and audit best. Baked into the gadget instructed, no longer tips.
Structured handoffs, no longer uncooked activates. Brokers do not obtain your unique request and interpret it freely. Each and every one begins from a structured temporary the upstream agent produced. So although one agent produces surprising output, the following is scoped by way of what it won — no longer a unfastened reinterpretation of what you initially requested.
Self-escalation when scope breaks. When an agent hits ambiguous necessities or conflicting constraints, it does not wager. It stops, prints precisely what resolution is wanted, and waits. 3 HITL varieties — GATE, CHECKPOINT, ESCALATION. Each and every one inflexible in structure: what stopped, what resolution you want to make, what it is aware of and does not.
The failure mode you are describing occurs when brokers have unfastened scope and no structured interfaces between them. The repair is not higher activates — it is making the interfaces between brokers as strict as serve as signatures.
What failure mode are you hitting in particular?
