intent-orchestration-contract

To safely route user intent to multiple agents, use a contract-first orchestration plan that transforms intent and context into structured tasks with validation, rather than relying on hardcoded agent names or ad-hoc string routing.

An OrchestrationPlan is a standardized JSON format that structures multi-agent tasks by defining execution strategy, expected outputs, dependencies, and fallback semantics to ensure validated and traceable agent routing.

To validate an LLM planner output, constrain the planner to generate JSON candidates and apply local schema validation plus capability, agent, and expected-output checks before any execution is permitted.

Yes, this orchestration contract supports group conversations and @mention routing by defining manual, mention, direct, and auto planning modes with clear routing priorities and audit trace requirements.

When multi-agent task planning fails validation checks, the system triggers bounded fallback modes with re-validation and returns a SafeError response to ensure execution halts safely without unpredictable behavior.

Yes, you can execute parallel agent tasks by defining an ordered_parallel strategy within the orchestration contract, which includes bounded fallback and retry semantics to handle execution failures safely.