Rensei is the enterprise OS for AI agent fleets. Vendor-neutral orchestration across any model, tracker, and SDLC. Open core. Governance for the humans accountable for what ships. Founded by a third-generation builder. 錬成: to forge, to temper, to refine.

A2A agents surface as MCP tools. MCP tools expose as A2A skills. Cedar policy fires in both directions regardless of which model submitted the request. Vendor neutrality as a structural property, not a claim.

Agent code that survives 30+ days in production gets more work. Reverted or rewritten code does not. Survival checkpoints at 7, 30, and 90 days feed Thompson Sampling posteriors. The router prefers agents whose code lasts. rensei.ai/docs/analytics/fac…

One orchestrator. Claude, Codex, Gemini, Devin. Each issue routed to the agent whose code survives longest in production. Cedar policy fires before any agent touches a task. The model changes. The governance does not.

Nightly 10-step smoke harness. If the daemon fails to install on a fresh OS, a Linear blocker is filed before anyone wakes up. That is what nightly CI looks like for on-prem agent infra.

Every MCP tool is now an A2A skill. Every A2A agent is now an MCP tool. Cedar fires at Layer 6 pre-verb in both directions. Protocol choice is no longer a lock-in decision. Platform v0.2.0.

Under 100ms. File-touch trigger to surfaced agent memory context, injected mid-session via Layer 6 hook bus. Not start-of-session only. Configurable per work type. REN-1184.

Integration breadth is only useful if governance stays in the path. Rensei has code for GitHub, Linear, Jira, Asana, Vercel, Slack, incident_io, WorkOS, SCIM, Claude, OpenAI/Codex, Gemini, Ollama, OpenCode, and multiple sandboxes. Cedar makes the list coherent.

Agent invocation is now a workflow node, not a side path. `agent.invoke` delegates into the same dispatch stage as the rest of Rensei: AgentCard discovery, pool eligibility, model-profile routing, Cedar-gated A2A dispatch, cost events. The control plane stays in the hot path.