# Guard your agent. Prove what it did.

Product / AgentGuard

See the whole run. Apply the right controls. Keep the evidence.

[Try AgentGuard](https://www.actaclad.com/contact)

## Banking

Prepare a payment to a new payee.

### Reason freely. Act within bounds.

- The customer request opens one trace.
- Customer context is protected before the model call.
- The model call stays connected to its context.
- Account checks run read-only.
- The payment tool requires approval. The action stays held.
- The response makes the held action clear.
- Output checks and control decisions stay with the trace.
- One run. One clear record.

### One run. One clear record.

- Payment action held
- Sensitive fields protected
- Response evaluated
- Decision evidence retained


## Healthcare

Show my care information and prepare a follow-up request.

### Useful context. Explicit boundaries.

- The patient request opens one trace.
- Scoped patient context passes through sensitive-data protection.
- The model receives the protected context.
- Health information is retrieved through a read-only tool.
- The request tool requires approval. No action is executed.
- The response passes through content-safety checks.
- Output evaluation and decisions are retained together.
- One run. One clear record.

### One run. One clear record.

- Sensitive fields protected
- Request action held
- Content safety checked
- Decision evidence retained


## Aviation

Review a disruption and prepare an operational change.

### Connect the systems. Hold the action.

- The operations request opens one trace.
- Flight, crew and gate context branch into the same run.
- Reasoning brings the three paths together.
- The operational tool reaches its policy boundary.
- APPROVAL_REQUIRED. The operational action stays HELD.
- The response reports the held action.
- The response is evaluated and its decisions retained.
- One run. One clear record.

### One run. One clear record.

- Operational action held
- Tool boundary enforced
- Response evaluated
- Decision evidence retained


Illustrative runs using configured controls. Actions remain held; no payment, care or operational change is executed.

## Set the controls. Guard the run.

The AgentGuard Controls screen brings model access, tool permissions and resource limits into one place.

[Inspect the product capture](https://www.actaclad.com/product/agent-policies.webp)

## Models, frameworks and protocols

OpenAI: verified SDK integrations.

Anthropic: verified SDK integrations.

LangChain: verified SDK integrations.

LangGraph: verified SDK integrations.

MCP: verified SDK integrations. Tool-call instrumentation for MCP clients and servers.

## Deployment

- SaaS
- BYOC
- VM

## One run. Every decision, connected.

AgentGuard captures each agent run as a trace: model calls, tool calls, latency and token usage.

AgentGuard guardrails detect and redact sensitive data such as PII, secrets and tokens in prompts and outputs.

Custom policies apply keyword or regex rules to inputs and outputs, scoped to selected agents.

AgentGuard checks model responses with output guardrails such as schema validation, hallucination and toxic-content checks.

AgentGuard tool_permission sets each tool an agent can call to ALLOW, READ_ONLY, APPROVAL_REQUIRED or BLOCK, with a per-tool call cap.

AgentGuard records each control decision as evidence.


## Your stack. Your deployment.

Built for multi-model, multi-framework AI environments. Deploy your way.


## From the first model call to the final action, AgentGuard keeps the run, the controls and the evidence connected.


[Contact ActaClad](https://www.actaclad.com/contact)