Project · Design a multi-agent context architecture

Take a 3-4 agent workflow (e.g. researcher → analyst → writer → reviewer) and design its context architecture: per-agent scopes, typed hand-offs, an orchestrator that keeps only summaries, provenance tagging to prevent contamination, and a context-flow trace. Implement it and demonstrate that no agent sees more than it should and the orchestrator's context stays bounded.

• A working multi-agent pipeline with explicit per-agent context scopes and typed hand-offs.
• An orchestrator that stores only summaries, plus provenance tagging on shared state.
• A context-flow trace that logs each agent's input/output and flags scope violations.

• No agent's window contains data outside its declared scope (demonstrated via the trace).
• Orchestrator context grows with summary count, not total work volume.
• Hand-offs are typed and validated; a dropped field is caught, not silent.