Product
Governed Agents
Coordinators and delegates
An agent can be one worker or a small org: a coordinator that owns the conversation, and delegates that own specialties. Who exists and who hands off to whom is declared in configuration, never improvised at runtime.
Skills
Skills are the agent's declared interfaces to the outside: interactive skills for conversation, and MCP-compatible connections for tools. A capability that isn't declared doesn't exist.
Instructions and memory
Instructions are managed documents, not throwaway prompts. Long-term memory carries what the agent learns across runs.
Operations are the only exit
However capable the model, the only way to change business data is a defined operation, and every operation passes the rule engine first.
The runtime
Agent work runs as isolated jobs on a dedicated machine pool. It scales with load, retries from feedback when a run fails, and keeps every workflow's data and execution separate.
Isolated jobs
Each run executes in its own job, so parallel work can't corrupt shared state.
Scales with load
The machine pool grows and shrinks with demand, and queues absorb the spikes.
Retries from feedback
A failed run isn't lost work: the agent picks up the feedback and tries again.
Rule Engine
Every write passes through the rule engine before it executes. There are exactly three outcomes: pass, wait for approval, or block. A write that fails the check doesn't happen.
Named operations
The only write paths are operations you've defined: named, typed, with declared failure modes. Undefined writes don't exist.
Pre-execution rules
Caps, separation of duties, state transitions, uniqueness. Rules are evaluated mechanically before commit, outside the model.
Approval checkpoints
The judgment calls you name wait for a person, and no agent can approve its own write.
Rules live in the definition
Rules are part of the versioned workflow definition, not a prompt. Every change lands with a diff and a date.