Runtime control plane for AI agents

Approve risky agent actions before they hit production.

AgentGuard traces every prompt, tool call, policy decision, and approval. Built for AI agencies and SaaS teams that need governance evidence without slowing down delivery.

Apply for design partner pilot Open live demo

5core governance views in the first demo

<30mtarget SDK integration for a monitored agent

€500design partner pilot with custom policies

Live control surface

A polished console for agents that can take real action.

The demo stays narrow on purpose: a support agent attempts a high-value Stripe refund, AgentGuard catches the risk, pauses execution, and records the evidence.

support-refund-agent / production

Views

Trace timeline

Support agent handling a damaged-order refund request.

policy triggered

Input received

"My order arrived broken. Refund me now."

00:00.000

Context retrieved

Order #A-10492, customer tier Gold, damaged delivery note, refund history clean.

00:00.034

Tool call requested

stripe.refund with { amount: 24900, customerId: "cus_123" }

high risk

Policy evaluated

Refunds over €100 require approval before execution.

paused

Awaiting decision

The requested refund is waiting in the approval queue.

pending

Risk events

Prioritized issues from live support-agent sessions.

3 open

approval required

High-value refund attempted

Agent requested stripe.refund for €249, exceeding the threshold.

blocked

Internal note in external reply

Generated response referenced private_comment, forbidden for customer messages.

warning

PII detected in output

External answer contained a full street address and was flagged for review.

observed

Cost spike

Model spend reached 3.8x baseline after repeated retrieval calls.

Approval queue

Risky tool calls pause until a human approves or rejects them.

pending

Issue refund for damaged order

The agent is paused before a high-value refund reaches Stripe.

waiting for review

Toolstripe.refund

Amount€249.00

Policyrefund_over_100_requires_approval

This action is paused. Review the trace and approve or reject before execution.

Policies

Simple rules decide whether actions are allowed, warned, blocked, or routed for approval.

Refunds over €100 require approval

Pause high-value refunds for human review.

when tool = stripe.refund and amount_gte = 100

Never expose internal notes

Block customer-facing output containing internal-only markers.

when output contains private_comment

Warn on PII in external messages

Flag names, addresses, account numbers, and sensitive identifiers.

when output contains sensitive_identifier

Unapproved tools are blocked

Allow only whitelisted tools for the current agent and environment.

when tool not in allowlist

Audit export

Evidence pack for client reviews, questionnaires, and internal governance meetings.

Live demo evidence summary

Traces1

Tool calls1

Approvals1

Blocked0

ready

Report includes policy decisions, approval timestamps, approver identity, final outcomes, risk events, cost summary, latency summary, and a detailed event appendix.

Positioning: evidence for governance and audit preparation, not a promise of legal compliance.

Control primitives

Runtime governance that agencies can explain to clients.

Keep the product surface focused: observe agent behavior, apply policy, pause risky actions, and export proof.

Agent tracing

Log prompts, retrieved context, model calls, cost, latency, and every tool attempt.

Policy enforcement

Allow, warn, block, or require approval based on rules clients can understand.

Human approvals

Pause risky actions before execution and record the reviewer, reason, and outcome.

Audit evidence

Export proof of what happened, which policies applied, and who approved actions.

Refund-agent use case

The agent is useful. The risky action is controlled.

The scenario is narrow on purpose: a support agent tries to issue a €249 refund. AgentGuard traces context, catches the risky tool call, applies policy, pauses execution, and records the human decision.

Agent gathers context

Customer tier, order history, damage report, refund history, and ticket metadata are logged.

Tool call is evaluated

The requested refund exceeds the approved autonomous-action threshold.

Human review happens

The queue shows the trace, policy, amount, and customer context before approval.

Evidence is exported

The client gets proof of oversight, policy enforcement, and final outcome.

Design partner pilot

Looking for 3 design partners.

Best fit: AI agencies deploying support, sales, or operations agents for mid-market clients. Pilot includes onboarding, custom policy setup, private support, and weekly risk review.

Start pilot conversation