Inside the runtime.

The connectors do the boring half of the job — credentials, retries, rate-limit politeness, schema translation — so the rest of the system can treat Stripe events, Gmail threads, GitHub issues, and PostHog feature flags as one data layer. Add a connector, every agent gets new sight; deprecate one, every agent silently stops reading. The primitives layer above is where signal-by-signal becomes story-by-customer: Lou's billing flag, Hank's quiet user, and Frankie's escalation collapse into one Person record because they were always the same person.

Above the primitives sits the agent harness — a playbook runtime that loads the right tools, applies guardrails, drafts work, and routes it through the approval queue. Each agent is configured, not coded. Playbooks are real artifacts you can read, version, fork, and pin to a customer segment. No prompt engineering passes the property line. The harness writes nothing back to your stack without a logged decision, which is the contract that turns "five AI coworkers" from a marketing line into something a CFO will sign.

The audit log is the foundation, not the afterthought. Every read, every draft, every approved action lands in an immutable record keyed to the customer, the playbook version, and the human who approved. SOC 2 readiness follows from the log; so does the ability to answer "why did this email get sent on Tuesday?" three months later. The whole runtime speaks MCP — the open protocol most of the ecosystem already speaks — which means the same connector library and primitives are useful inside Claude Desktop, Cursor, and any tool that follows the standard. Open at the edges, opinionated in the middle.