Multi-agent guidance

Distilled from the published multi-agent research-system engineering (source S3 on the Citations page), applied to how agent definitions and run contracts are written here:

1. Delegate with a full specification. Every dispatched worker gets an objective, an output format, guidance on tools and sources, and clear task boundaries. Vague delegation produced duplicated and misaligned work in production.
2. Scale effort to the job. Simple fact-finding warrants one worker and a handful of tool calls; a direct comparison warrants a few workers with ten to fifteen calls each. These rules live in the run contract, not in anyone's head.
3. Budget tokens deliberately. Token spend explained most of the performance variance in the published evaluation, and multi-agent runs cost roughly fifteen times a chat. The dispatch rules at the gateway exist to spend that only where the outcome justifies it.
4. Parallelize tool calls. Parallel calling cut research time by up to ninety percent on complex queries; workers fan out concurrently wherever the dependency graph allows.
5. Let the model repair its own prompts. Given a failing prompt and the failure mode, the model diagnoses and improves it; the published result was a forty percent reduction in task time. This runs inside the ADLC, and every improvement ships as a versioned change.
6. Judge end states, not steps. Evaluations score whether the correct final state was reached, with an LLM judge scoring factual accuracy, citation accuracy, completeness, source quality and tool efficiency; people still test by hand, because they find what judges miss.
7. Resume, never restart. Long runs summarize completed phases into external memory and continue from checkpoints; deploys shift traffic gradually so running agents are never disrupted. Both are harness inventory items in Environment & Operations.