Anuclei Expands Multisynapse From Agent Certification to Conversation Governance

Summary

Anuclei says its latest Multisynapse release extends certification from an individual AI agent to the conversations, agent compositions, and evidence surrounding its actions. Each run now receives a deterministic diagnosis drawn from platform records, identifying policy refusals, unavailable tools, empty knowledge lookups, context trimming, outdated revisions, and failed judge results. Optional narratives are generated only from those findings and are marked stale if the underlying record changes. Online evaluation can score a complete multi-turn conversation, using prior turns and tool results as context, and any conversation can be promoted into a golden regression example. The platform also governs coding agents operating through its MCP server: calls are attributed to a specific client, checked against policy, subject to kill switches and budgets, and shown in a client-filterable feed. Coding-agent tokens expire by default. Other changes record scorer provenance, prevent unintended historical backfills, measure judges under production settings, enforce model allowlists across agents and evaluation tools, and attach notes to saved versions. A new swarm view compares declared agent handoffs with observed ones, including paths that were used but never certified. Anuclei says these changes point toward certifying agent compositions and conversations rather than isolated agent nodes; behavioral identity for an entire composition and evidence-based trust decay remain roadmap ideas.