Skip to main content

Agent Governance

Clevername provides continuous enforcement of what AI agents are allowed to do. From the moment an agent is governed through every action it takes, governance covers the full lifecycle: trust scoring, drift detection, integration bindings, and fleet monitoring.

Key Concepts

Agent lifecycle

Every agent follows a defined lifecycle: draftagent review approvedactive. When the Agent Review is enabled, agents cannot transition from draft to active without passing a structured security review and receiving enough governance votes. Once active, the agent operates under its compiled guardrail profile, which is enforced at every tool call.

Continuous enforcement

Clevername does not just govern agents once and forget them. The governance system is continuous: every action an agent takes is validated against its guardrail profile, trust score, and integration bindings in real time. If the agent drifts from its governed mandate, it is detected and alerted by default, and blocked once enforcement is enabled.

What makes this different from IAM

Identity and access management (IAM) asks “who are you?” once at the door. Content scanners inspect data after the fact. Clevername enforces the governed mandateof each agent continuously — the gap between what was governed and what agents actually do.

Trust Scoring

Every agent has a dynamic trust score from 0 to 100 that is continuously updated based on its behavior. The score is computed from six weighted dimensions:

DimensionWeightWhat it measures
Scanner Compliance30%How often the agent's content passes CleverGuard scans without findings
Tool Compliance20%Whether the agent stays within its governed tool scope
Budget Compliance15%Whether the agent stays within its cost budget
Behavioral Stability15%Consistency of the agent's behavior patterns over time
Human Sign-off Compliance10%Whether high-risk actions are properly signed off before execution
Session Health10%Clean session completion rate without errors or timeouts

Trust scores drive enforcement tiersthat automatically adjust the agent's autonomy level:

Full Autonomy80-100

Agent operates independently. No approval required for standard actions.

Supervised50-79

Agent requires human sign-off for sensitive actions. Rate-limited to 80%.

Restricted20-49

Agent requires human sign-off for all actions. Rate-limited to 30%.

Suspended0-19

Agent is frozen. No actions permitted until score recovers or admin intervenes.

Note
Trust scores recover over time when agents behave well. Each clean session, compliant tool call, and clean scan result adds a small positive increment to the relevant dimension.

Drift Detection

After an agent is governed through Agent Review, Clevername computes a version hashof its configuration: model, system prompt, expertise, environment, and integration bindings. If any of these change after governance, the hash no longer matches and the agent's status transitions to resubmit_required.

Drift detection operates at three levels:

  • Configuration drift— the agent's model, system prompt, tools, or bindings were changed after governance. Blocks deployment until re-review.
  • Tool drift— the agent attempts to call a tool that is not in its governed allowed_tools list. Blocked at runtime with error code -32403.
  • Behavioral drift— the agent's runtime behavior deviates from its review baseline (e.g., calling tools in unusual patterns, elevated error rates). Flagged in the SOC Console and may trigger automatic trust score reduction.
Drift detection dashboard showing a timeline of agent configuration changes, with flagged drift events highlighted in red
The drift timeline shows every configuration change and runtime deviation, with automatic alerts for critical drift.
Important
High-risk drift events (like attempting to call blocked tools) immediately reduce the agent's trust score by 10-15 points. Repeated violations can push the agent into the Restricted or Suspended tier.

Integration Bindings

Integration bindings define exactly which external services an agent is allowed to connect to. When an agent is submitted for Agent Review, its bindings are recorded as part of the governance review. At runtime, the gateway checks every tool call against the binding list before execution.

Bindings can include:

  • MCP servers — which MCP servers the agent can use
  • Tool allowlists — specific tools within each server
  • API endpoints — external HTTP endpoints the agent can call
  • Data sources — databases, file stores, or APIs the agent can read from

If an agent attempts to call a tool from a server that is not in its bindings, the call is blocked and logged as a drift event. The binding fingerprint is included in the version hash, so adding or removing bindings after governance triggers the resubmit_required state.

Tip
When designing agent bindings, follow the principle of least privilege: only bind the specific servers and tools the agent needs for its defined purpose.

Fleet Monitoring

For organizations running multiple agents, Clevername provides fleet-level monitoring across several dashboard views:

SOC Console/dashboard/security/socReal-time view of all active agent sessions, spend tracking, security alerts, and DLP findings.
Drift Monitoring/dashboard/security/driftFleet-wide drift event timeline with severity filtering, agent grouping, and trend analysis.
Agent Health/dashboard/ai-companyComposite health scores (0-100) combining trust, drift, errors, budget, and scan results per agent.
Compliance Export/dashboard/security/complianceGenerate SOC 2 / HIPAA evidence exports covering agent activity, governance outcomes, and incident responses.
Fleet monitoring dashboard showing agent health scores in a grid view, with color-coded status indicators (green for healthy, yellow for degraded, red for critical)
The fleet monitor provides at-a-glance health for every agent in your organization.

Fleet monitoring also supports contagion detection: if one agent's trust score drops critically (below 30), agents that share the same tools or data sources receive a small preventive penalty (-5 points per dimension) to flag potential cross-agent risks. This cascade is capped to prevent runaway score collapse.

Policy Tiers at Dispatch

When an agent is dispatched to handle a work item, its guardrail profile is loaded and a trust tier is assigned based on the profile's risk score:

Risk ScoreTrust TierEffect
0 – 30TrustedFull access. Minimal runtime overhead.
31 – 60StandardNormal enforcement. All scans active.
61 – 100Elevated ReviewEnhanced scanning. High-risk actions require human sign-off.
No profileUntrustedMaximum restrictions. Agent must complete Agent Review.
Note
Risk scores are clamped to the range 0–100 and validated on profile creation. Negative scores cannot be used to bypass trust tier enforcement.