Introduction to governed agents
Model an agent as a bounded decision-and-action system with explicit authority, tools, memory, and evidence.
Agents cross the decision-action boundary
An agent observes context, plans or selects steps, invokes tools, changes external state, uses memory, and adapts to results. Governance must constrain the complete loop, not only the model prompt.
Define the agent's purpose, prohibited goals, initiating actors, affected resources, data, tools, side effects, autonomy, duration, budgets, memory, external dependencies, and stop conditions.
Principles
- Treat model output as an untrusted proposal, not authorization.
- Allowlist tools with typed contracts and bounded effects.
- Preserve actor, agent, workload, and tool identity.
- Authorize each material action at execution time.
- Require approval where consequence or uncertainty demands it.
- Isolate execution and cap time, spend, steps, data, and network reach.
- Produce attributable run receipts and test recovery.
Module outcome
You will create an authority boundary, tool registry, approval and confinement policy, run-receipt schema, and evaluation/incident plan.
SAI-280 is roadmap material: the conceptual design is public, while executable labs remain unreleased pending reproducibility and safety validation.