Automatically allowed
- Requested action
- Add an internal note to an authorised record
- Example rule
- Internal notes to authorised records are permitted
- Decision
ALLOW- What follows
- The connected service proceeds automatically.
Let your AI agents work. Keep control of their actions.
HaltState automatically checks connected tool and API actions against your policies before execution. Let permitted work proceed, block prohibited operations, and hold exceptions for the approval your policies require.
AI-assisted policy setup. Deterministic enforcement. Signed audit evidence.
Control the action, not just the conversation
An agent can propose a useful task and still request an operation outside its authority. HaltState evaluates the connected action at the point where your service would execute it.
Automatically blocked
- Requested action
- Export customer data to an unapproved destination
- Example rule
- Export destination must be approved
- Decision
DENY- What follows
- The connected action is blocked.
Held for review
- Requested action
- Change a protected production setting
- Example rule
- This change requires operator authorisation
- Decision
APPROVAL_REQUIRED- What follows
- The action is held for review.
QUARANTINE: A configured policy can stop further governed actions for an affected agent or session.
Put your rules where actions execute
Define the authority
Identify the tools and APIs your agents use and the context needed to evaluate each action. Use AI assistance to draft policies, then review and activate them.
Enforce automatically
The deterministic engine evaluates the active rules and supplied action context. Routine permitted work proceeds without a person clicking Approve.
Keep human control where it matters
Hold policy-defined exceptions for the required reviewer. Changing the policy and authorising one particular action remain separate responsibilities.
Keep reviewable evidence
Use Signed Proof Packs to inspect recorded decisions, approvals and reported outcomes.
Fit control into your existing stack
Your application calls HaltState at the connected tool or API boundary before it executes the operation. Your service follows the returned decision and reports the outcome. Existing identity, network and host protections remain part of your stack.
- Agent request
- Connected service
- HaltState policy decision
- Permitted execution or hold/block
- Reported outcome
Explore the engineering context: AI agent kill-switch design patterns and why model guardrails need an action boundary.
Evaluate one workflow with your team
Choose a consequential action, define what is permitted and what needs review, then connect it to policy enforcement. Use the decision and evidence path to assess how it fits your operating model.
Which system changes? Who owns the policy? Which actions need a reviewer? What evidence does the team need afterwards?
Ready to implement? Integrate an agent.
Buyer questions
Does someone have to approve every action?
No. Policies automatically permit or block the actions they can resolve. Human review applies to the cases configured to require it.
How does this relate to model guardrails?
HaltState adds policy decisions at the connected action boundary. It evaluates the operation your service is about to perform, alongside the model and infrastructure controls you already use.
What does the integration involve?
Connect the relevant service or tool calls through the existing SDK/API path, supply the context your rules need, follow the decision and report the outcome. The integration guide gives developers the next steps.
What can we inspect afterwards?
The product's evidence views support review of recorded decisions, approvals and reported outcomes. The Signed Proof Packs page explains the available verification checks.