Runtime Governance

Policy at the action boundary.

Runtime governance at the action boundary — wherever AI agents execute — not post-hoc API monitoring. Pre-execution interception, intelligent triage, and cryptographic tool integrity, in one control plane.

See governance in action →Guard module
Enforcement modes

Allow · deny · ask — every action.

  • Allow

    Action proceeds; decision logged to Ledger.

  • Deny

    Action blocked at the boundary—no proxy bypass.

  • Ask

    Step-up approval for shell, MCP, or sensitive file ops.

Policy-as-code

Versioned YAML. Reviewable in CI.

rules:
  - id: block-prod-shell
    match:
      action: shell.exec
      env: production
    effect: deny
    reason: "Destructive shell in production requires approval"

  - id: mcp-egress
    match:
      action: mcp.call
      destination: external
    effect: ask
    approvers: [security-oncall]
Pre-Execution Reasoning Drift Guard

Catch the bad action before the tool is called.

Most controls catch the tool call after the agent has already decided. ComplyAI watches the decision form, against the original objective, and ends the reasoning chain when it drifts — before any tool is invoked.

  1. 01

    Objective anchor

    System prompt and policy intent locked at session start.

  2. 02

    Continuous drift score

    Every reasoning step scored against the anchor on three axes — semantic, coordination, and behavioural drift.

  3. 03

    Pre-execution terminate

    If ASI breaches threshold, the chain ends before any tool is invoked. Full reasoning trace logged.

HITL Smart Triage · Inside Govern

100,000 actions in. 800 land on a human — by design.

Approval queues only work if a human actually reads them. At fleet scale, they don't. ComplyAI's triage layer pre-audits the queue and surfaces the small share that genuinely needs a decision.

  1. 01Agent actions per day

    100,000

    Across your full agent fleet — tool calls, file ops, MCP requests, shell, API.

  2. 02Policy engine — pre-execution

    ~96,000 allowed inline

    Routine, in-scope, identity-clean actions allowed at the action boundary in <10ms.

  3. 03HITL Smart Triage

    ~3,200 auto-approved

    Constrained companion model summarises action context, applies precedent, auto-approves low-risk.

  4. 04Genuine human review queue

    ~800 actions

    Plain-English summaries. Single click. Denials auto-update agent system prompt.

  5. 05Blocked or escalated

    ~80 actions

    Drift detected, signature mismatch, blast-radius breach, or policy deny — full audit chain captured.

Early access

Your agents are already taking actions. Governance shouldn't be an afterthought.

ComplyAI is in early access. We're working directly with security and engineering teams to deploy, configure, and demonstrate value in their environment — in a single session.

Request early access See the platform