The platform

One checkpoint, three places it has to hold.

These aren't three separate bets. Governing what agents do produces the evidence regulators want, and both are the prerequisite for trusting machines that move on their own.

Sequence

Each layer is usable alone, and worth more when the one below it is already running.

The foundationShipping from 2026

AgentGuard

A checkpoint in front of every action your agents try to take.

AgentGuard stands on the execution path. Before a tool runs, the request is attributed to a specific agent, matched against rules you can read, and returned as one of four outcomes — run it, trim it, ask a human, or stop it. The model never gets a vote.

curiousdevs platform · agentguard
Preview

184,402

Actions reviewed today

312

Stopped before execution

27

Sent for human approval

1,148

Rules in force

Verdict time

under 10 ms

Sits at

the tool call

Record

append-only

  • A distinct identity and permission budget for each agent
  • Plain, versioned rules instead of instructions buried in prompts
  • Injection and goal-drift detection on every call, before it executes
  • Approval queues for high-value actions, with timeouts that fail closed
  • Signed record of every request and the reason behind its verdict
  • Verification of the tools and connectors agents are allowed to reach
  • Open-core: self-host free, no seat limits — managed cloud is optional
Walk through AgentGuard

Runtime execution graph

Five modules, one path to a verdict.

Every tool-call walks the same graph, in the same order. Nothing skips a stage, and nothing executes before the checkpoint returns a verdict.

IDENTITYAgent IAMENFORCEPolicy EngineINSPECTThreat DetectionRECORDAudit & ForensicsREGISTRYMCP SecurityVerdict

Enforcement capabilities

Complete surface coverage.

Policies are plain YAML, versioned in git — reviewable, testable, auditable.

  • Argument limits, spend caps, rate limits, time windows
  • Human-in-the-loop escalation for sensitive actions
  • Dry-run mode against historical traffic before enforcement
Escalation Request
Agentsupport-agent
Actiontool.refund
Amount₹50,000
RiskHIGH · 87/100
ApproveDeny

Let something decide for you. Just not unsupervised.

Bring one workflow you are nervous about. We will show you exactly where the checkpoint goes and what it would have stopped last month.

See the platform

Questions

The things people ask first

Short answers on how this runs, what it costs you in speed, and what you can hand to an auditor.

A decision checkpoint for software and machines that act on their own — it approves, trims, escalates or refuses each action, and keeps a signed record of why.