Warden is an enterprise AI agent studio where every agent is governed by default. You give an agent instructions and grant it tools over MCP. Reads run on their own; anything that moves money or changes state is held for a human. Every step lands on a tamper-evident record.
A lead agent hands tasks to member agents with one governed tool, delegate. Members keep their own grants and approvals, the lead never inherits their tools, one budget covers the whole team, and a member’s approval pauses everyone until a human decides. Every hand-off is on the record.
Agents are probabilistic, so Warden measures rates. Mark what goes wrong in real conversations, turn them into cases, score with code checks first and an LLM judge only where judgment is needed, and compare every experiment against its baseline check by check. Eval runs only ever read: anything above a read is recorded as held, never executed.
Every tool is risk-classified the moment a server connects. A policy engine adds spend caps, rate limits, and time windows that decide allow, gate, or deny per call.
High-risk actions pause for approval, with a full diff view for proposed code changes. Nothing irreversible runs without a user saying yes.
Every step is hash-chained, so any edit, deletion, or reorder is detectable. The record of what an agent did outlives the agent itself.
Estimate a run before it starts, watch the cost meter live as it works, and cap spend per run. Visible, then enforceable.
A first-party server catalog plus live discovery from the official MCP Registry, official vendor servers only, never third-party middlemen.
Each run is an OpenTelemetry trace, each step a span. Sensitive fields are redacted before export to Datadog or Grafana.