Govern an agentic workflow
When AI stops answering and starts acting, calling tools, moving data, making changes, design-time assessment is no longer enough. This guide puts an agentic workflow under enforced runtime governance.
When to use this
Section titled “When to use this”You have, or are building, an agent that takes action: it calls tools and APIs, retrieves and writes data, or runs multi-step workflows, whether on Gamut Claw or your own framework.
What you will produce
Section titled “What you will produce”A registered, ATF-assessed agent designed to act only through Gateway-enforced authority, with provider credentials outside the agent and request-level runtime evidence.
- Register the agent. Add it to the Agentic CISO agent register with a human owner and a security owner. An unregistered agent is blocked from acting at all.
- Set its autonomy. Assign an ATF level (Intern through Principal). Gateway enforces a different action boundary at each level, from read-only to strategic autonomy.
- Score its capability risk. Run an ACRS assessment to set how tightly to govern, and model the threats its capabilities introduce with MAESTRO.
- Authorise its tools. Grant tool permissions in Agentic CISO and confirm the matching governed connectors exist. An agent can use a tool only when both layers agree.
- Define Runtime Access Policies. Bind each required authority to the exact agent, governed connection, action, resource, data classification, environment, purpose and time boundary. Validate the policy, review its impact and submit the immutable version for independent approval.
- Configure approval gates. Require human approval for sensitive or mutating actions (external calls, financial actions, code changes). Gateway enforces them on every request.
- Independently approve runtime authority. A workspace-scoped Runtime Policy Approver reviews the submitted policy and either publishes it or rejects it with a reason. Approval requires the reviewer’s own MFA-protected step-up.
- Choose the runtime. Run it on Gamut Claw or, for an external framework, the BYO runtime. Either way the rule holds: think anywhere, act through Gateway. Credentials live on Gateway, never with the agent.
- Simulate, then watch. Run a Gateway simulation to see what Gateway would decide and close gaps before going live, then rely on the hash-chained runtime evidence and the audit log for ongoing oversight.
Go-live evidence and gates
Section titled “Go-live evidence and gates”- Authenticated agent identity, owners and purpose boundary.
- ATF target, ACRS risk and MAESTRO threat treatment.
- Exact Tool Permissions, governed connections and data flows.
- Independently approved Runtime Access Policies with expiry and review.
- Allow, deny, approval, timeout, replay, failure and containment tests.
- Monitoring ownership, incident playbook and kill/suspend procedure.
- Residual risk decision and material change triggers.
Readiness is not standing authority. Gateway revalidates the exact request immediately before execution, and missing or stale context must stop the action.
Modules and frameworks involved
Section titled “Modules and frameworks involved”Agentic CISO, Runtime Access Policies, Gateway, Claw, BYO runtime, ATF, ACRS and MAESTRO.
- Agentic stack overview: the full architecture behind this guide.
- Govern a GenAI chatbot: for assistants that only answer.
- Scenario guides: the full set.