Loading documentation
Autonomous systems / 02
Control / Human-led
A governed agent is an operational role with bounded context, permitted tools, measurable confidence, escalation rules, and an accountable human owner. Language instructions support the role; policy controls it.
Every agent begins with a narrow operational purpose.
Reduce both what the agent can know and what it can change.
Provide only mission-relevant, provenance-marked context within the approved data boundary.
Allow only named tools with target, parameter, rate, time, and environment limits.
Validate structured proposals before they become commands or external changes.
Expire authority automatically when the mission window or operating condition changes.
Multi-agent operation works best when responsibilities do not blur.
Low confidence changes the route; it never disappears inside a narrative response.
Attach confidence, source freshness, assumptions, and unresolved conflicts to each proposal.
Escalate when confidence, evidence quality, policy, or timing leaves the approved envelope.
Pause or degrade when required human authority is unavailable.
Record the proposal, policy evaluation, decision authority, and final outcome together.