Policy, authorization, audit and traceability for the teams responsible for keeping agent adoption within the organization's boundaries.
When each department secures its own agents, the organization ends up with as many security models as there are frameworks in use. A requirement approved once shouldn't need a separate implementation for every technology.
Policy enforced at the boundary, before the call, instead of explained afterwards.
Define once. Enforce everywhere.
Enforce security requirements centrally, not team by team.
Define which agents can reach which models, tools, data and systems.
Maintain traceability over configuration changes and runs.
Understand how an agent evolved and who changed it.
Role-based access, tenant isolation and data encrypted per data subject.
Identity per version, artifact digest, declared purpose and approval.
Declared tools and data; credentials from the vault, scoped and just in time.
Pause and approval are language effects, ask and approve, written in the code.
Guardrails, budget, token cap and policy, enforced at run time.
A hash-chained append-only trail, with a decision id and every external effect.
Fail-closed tool contracts, limits and continuous evaluation.
Pause, step-through, breakpoint and a kill switch that cancels in-flight runs.
Agent passport: one record per agent and environment, each answer marked declared, enforced or not covered. Deployment to production is blocked if the passport does not allow it.
Three ways to deploy: shared cloud, tenant isolation, or dedicated in the institution's own account, with its keys and a role it can revoke.