agent-context-engineering
Agent context engineering
Use the local reference implementation when available. Start with swarm-context-commander demo and inspect the resulting bundle and worker receipt. For a request's own documents, use swarm-context-mcp with compile_agent_context; its records are inline and untrusted. Set a token budget and compare selected source digests against expected sources. Do not treat a supplied tenant ID as authentication.
For runtime planning, use classify_agent_runtime to compare serverless, container, and microVM placement. That tool classifies a task; it does not provision or run workers. For capacity claims, run swarm-context-commander bench-fleet and report the host, counts, measured timings, and excluded workload dimensions. Logical-agent registration is not active-model concurrency.
When adapting this design to another system, keep the retrieval baseline and policy fixed while comparing new GraphRAG, vector, or compression adapters. Measure recall, contamination, cost per accepted task, p95 latency, and task recovery independently. Follow the benchmark protocol for definitions. Obtain explicit authorization before using customer data or enabling network integrations.