agentarch
A retrieval-grounded multi-agent system served over MCP. Every model, retrieval and clock call is journalled, so a run replays byte for byte offline, and a gate fails any claim its evidence does not support.
- Year
- 2026
- Kind
- Agents & RAG · Python
- Status
- Shipped
A planner, retrieval workers, a synthesizer and a critic reach the outside world through one effect boundary that writes a trace. A recorded run replays with no corpus, index or network, and replay refuses to go green once the code's behaviour changes. The system is an MCP server, and when its own retrieval is served over MCP the trace is byte-identical.
Retrieval was measured before anything generated. BM25 over 100-word chunks reaches 0.629 recall at a 2,000-word budget on 56 queries. Adding a dense arm and fusion gained 0.031, with a 95% interval of 0.000 to 0.076, so hybrid search was not shown to help. The grounding gate catches five injected lies and flags none of 13,899 correctly quoted claims.
No offline judge was fit to check entailment. Every candidate scored 0 of 198 on claims whose only edit is a negation, so that tier was not built. The offline agents are deterministic stand-ins, not a model, and the query labels were produced with a language model and never independently checked.
