Runtime governance: controls that adapt as agent behavior changes.

CompFly turns simulation findings into recommended runtime controls. Teams review the recommendations in plain English, tune the boundaries, and apply them before agents call models, tools, data, or other agents.

Where CompFly sits, on the request path.

CompFly sits inline with one line of code. The decision is made before anything reaches a model, a tool, another agent, or your data.

Your agent

Any framework, instrumented with one line of code.

CompFly
InterceptDecideEnforce

Every call decided before execution, against the controls designed for the agent.

Destinations

Models · Tools · Agents (A2A) · Data

example decisions
denied · unsanctioned modelheld · pending human approvaldenied · sensitive-data egress

Controls composed per agent, recomputed per request.

Org-wide baseline controls and per-agent controls combine into one set of effective controls for each agent, evaluated in priority order on every call.

Org-wide baseline

The controls every agent inherits. Set once, applied everywhere.

Per-agent controls

Specific to one agent: what it does, what it touches, how much it is trusted to do alone.

Effective controls

The exact combined set that runs on the agent, evaluated in priority order on every call.

The Autonomy Envelope: the exact set that will run, before you publish.

The Autonomy Envelope shows the combined set that will run on an agent, with each control’s source and action, before you publish. Nothing changes for a running agent until you publish to enforcement.

Every decision is recomputed on each call against who is calling, data sensitivity, destination, and how independently the agent is acting. The same agent gets tighter enforcement the moment the context turns risky, with no redeploy and no rule rewrite.

autonomy envelope · illustrative preview
  • baselineUnsanctioned model or tool calldeny
  • baselineCredential detected in tool argumentsdeny
  • per-agentEmail send to external domainapproval
  • per-agentTool call outside approved behaviorwarn

Define controls in plain language.

Controls Assist is the AI assistant built into CompFly. It recommends runtime enforcement controls with full context: what Agent Simulation found when it attacked the agent, what the behavioral baseline shows the agent actually does, and where its risk tier sits.

Pick the ones you want, or describe the control you need in plain language and it drafts the definition and refines it with you. You review and approve before anything enforces. It works from your own agents and controls, so its answers reflect your environment, not a generic manual.

controls assist · illustrative
you

Don’t let this agent email anyone outside the company without a human sign-off.

assist

Drafted a control:

email.send · external domain → require human approval

Review and approve before it enforces.

Stop chasing evidence for every AI agent.

CompFly continuously tests required controls, surfaces failures, and produces the evidence security teams, auditors, and regulators need.

Discovery & Visibility

Agent inventory, ownership and lifecycle status

Discovery & Topology
Decision & Control

Verifiable agent identity and credential lifecycle

DID / On-boarding
Grounding & Hallucination Resistance

Source grounding and anti-fabrication controls

Safety EvalsFunctional EvalsAdaptive Controls
Reliability

Risk taxonomy, acceptance criteria and test governance

Functional EvalsSafety EvalsSecurity Evals
Adversarial Security

Adversarial threat taxonomy and red-team program

Security Evals
Efficiency & Usage

Cost, token, latency, compute and action telemetry

Gateway / Observability

Every run compiles into a reproducible assurance report: runtime conformance, per-control pass rates with confidence intervals, a residual risk register, and framework mapping (AIUC-1, OWASP LLM Top 10, NIST AI RMF, MITRE ATLAS, ISO/IEC 42001, SOX).

A per-agent definition of normal, human-approved.

Sanctioning decides what an agent may use. Behavioral control governs how the approved agent behaves once running. It judges what the agent does at run time, not what the model says it will do. Static guardrails catch what a human named in advance; the events that hurt are the ones no rule named.

Intended profile

What the agent is sanctioned and configured to do: the approved tools, the operations, and the argument shapes it uses.

Observed profile

What the agent actually did, built from its allowed activity: every tool call, operation, and argument shape used in production. Anything observed but not intended is an anomaly.

Anomalies are deny-by-default, banded by risk.

The agent starts in a learning phase where nothing is ever blocked, so bringing an agent under behavioral governance never interrupts it mid-run. An administrator reviews each anomaly and either confirms it anomalous or marks it known with a justification (every review audited). Approving the baseline graduates the agent from Learning to Enforcing.

Learning Review Approve Enforcing

Only sanctioned models, tools, and MCP servers run.

Layered, broad to specific, so one decision at the right level covers everything beneath it.

Safety nets catch risky models even from sanctioned providers.
Per-model & per-tool
dept / env
Most specific
Domain
host / wildcard
Provider
whole provider
Pending-review
fallback
Broadest

Most specific match decides · covers everything beneath

Every agent carries its own cryptographic identity.

Every agent gets its own decentralized identifier (DID) and signing key, not a shared API key and not an IP address.

Authentication

One agent cannot impersonate another, every call is verified against the agent’s own key.

Least privilege

Controls attach to identity, starting from default-deny; an agent holds exactly what it was granted.

Attribution

Every recorded decision is tied to a specific agent, not a shared key.

Non-repudiation

Signed calls cannot be waved away; the record shows which agent did what.

A person in the loop, only where it matters.

Human approvals for actions too consequential for an agent alone: handling sensitive transactions, deleting data, sending email outside the org.

01

Denied on the wire

The call is denied immediately, so nothing hangs open; the agent surfaces “approval pending.”
02

A person approves with context

The agent, the tool, the destination, and the signals that triggered the hold, all in front of the approver.
03

Scoped, expiring override

The approval is cached; the next matching call proceeds as a recorded human-approved override, scoped and expiring.

Approvals arrive where teams work, including Slack. Holding too much trains approvers to rubber-stamp, so approvals are reserved for what matters.

Last resort: kill switch

An immediate, full stop on an agent, issued from CompFly. It overrides every protection mode and cannot be bypassed in code. When an agent needs to stop, it stops.

Secure, resilient, cost-aware access to every model.

Three modes of adoption, from advisory to fully managed. Start where your risk tolerance is today and move up without touching agent code.

Check

The agent calls the model directly; CompFly returns an advisory decision off the request path.

Passthrough

Calls route through the gateway with the agent’s own key; decisions are enforced server-side.

Gateway

CompFly-managed encrypted credentials are injected at call time; agents never hold the raw secret.

set org-wide · override per agent · switch modes with no code change

Resilience and cost.

Failover chains

Route across providers on provider errors so one outage does not take the agent down.

Key pools

Spread load across keys and cool down throttled ones automatically.

Canary routing

Split traffic between models for safe upgrades.

Cost visibility

Spend per agent, per model, per provider.

Providers

Amazon BedrockGoogle VertexOpenAIMicrosoft FoundryAnthropicYour own OpenAI-spec endpoints

Watch the control plane deny an out-of-policy call.

A live session on your stack: an agent attempts an out-of-policy call and the control plane denies it before execution.