Open Control Stack by Cognous | Runtime governance for AI agents
Apache 2.0 · Open specification · Reference implementation
Open Control Stack · Published by Cognous

Runtime governance for AI agents.

AI agents have started to act — calling tools, moving data, and changing enterprise systems. Open Control Stack authorizes what an agent may do, enforces policy before it runs, and turns every run into portable evidence.

Explore the stack View on GitHub
governed-run · lifecycle
Open, vendor-neutral, and free to adopt — no vendor agreement required.
Apache 2.0 Vendor-neutral JSON Schemas Conformance tests Reference implementation
Why this exists

AI agents need more than logs.

As AI agents begin using tools and triggering workflows, enterprises need to answer practical questions before and after each run: What was this agent allowed to propose? What did it actually propose? What did policy allow, block, or escalate? What did the agent rely on? Can the run be reconstructed later?

Traditional logs, dashboards, and policy documents are not enough by themselves. Governed AI-agent deployment requires structured runtime records and portable evidence artifacts.

One stack, four components

Declare, control, replay, and evidence AI-agent behavior.

Four open-source reference components, each mapping to one stage of the governed lifecycle. Use one, or run them in sequence.

Agent Action Manifest

Declare what an agent may propose.

A lightweight manifest format for documenting an agent's tools, action types, authority requirements, review posture, reliance requirements, payload policies, and redaction hints before runtime deployment.

Read the specification
Tool and action-type inventory
Authority requirements per action
Human review posture
Payload policy and redaction hints
Agent Control Plane

Control proposed actions at runtime.

A minimal runtime control layer that records action proposals, evaluates policy decisions, captures blocked actions, records authority and reliance, and generates replayable run records before agents affect enterprise systems.

Read the specification
Policy evaluation before execution
Allow, block, or escalate decisions
Authority and reliance capture
Replayable run records
Agent Replay Bundle

Package runs for reconstruction.

A portable replay-bundle format capturing the task frame, action proposals, policy decisions, evaluation traces, blocked actions, authority records, reliance records, validation reports, redaction metadata, and signature metadata.

Read the specification
Task frame and full proposal trace
Blocked-action records
Redaction metadata for safe export
Signature metadata for export integrity
Agent Governance Evidence Pack

Translate records into governance evidence.

A business-facing evidence-pack format summarizing agent purpose, deployment context, tools, actions, authority model, policy controls, blocked actions, reliance, replay bundles, validation results, redacted exports, risks, and review records.

Read the specification
Business-readable run summaries
Authority model and policy controls
Risk and review records
Redacted exports for external review
How it works

One path from intent to evidence.

Every governed action follows the same runtime path — proposed by the agent, authorized, evaluated against policy, executed, and recorded.

Agent
Proposes an action
Authority
Who authorized it
Policy
Allow, block, escalate
Execution
Acts on the system
Evidence
Recorded and provable
Who it is for

Built for teams moving agents into production.

The same artifacts serve engineering and governance — one record, read four different ways.

Enterprise AI teams

Move agents into production with governance built in, not bolted on.

Platform engineers

Add a runtime control layer without adopting a whole framework.

Security & governance

Review what an agent did — and prove it — in business-readable form.

Application developers

Declare an agent's actions once; get authorization, records, and replay.

What the stack helps teams answer

Concrete questions, before and after every run.

What actions can this agent propose?
Which tools and systems can it access?
Which actions require authority or human review?
What did the agent propose during a run?
What was allowed, blocked, or escalated?
What sources did the agent rely on?
Can the run be replayed or reconstructed?
Can sensitive fields be redacted before review?
Can replay artifacts be signed for export integrity?
Can governance teams review evidence in business-readable form?
Scope and boundaries

A reference stack, not a platform.

The Cognous Open Control Stack is intentionally narrow. It is a public reference layer that defines practical artifacts for governed AI-agent deployment — nothing more.

Not an agent framework
Not a model runtime
Not a hosted dashboard
Not a complete compliance platform
Not ODES — the stack may produce evidence that supports ODES-style records in future profiles, but it does not define ODES conformance.

Start with the open-source stack.

Open, vendor-neutral infrastructure for governing AI agents at runtime. Adopt one component, or run the whole stack.

View on GitHub Talk to Cognous