Agent Control Plane | Runtime control and replay for AI agents
Control/Agent Control Plane

The control plane for AI agents.

Agent Control Plane authorizes, records, explains, and replays agent behavior before AI systems affect enterprise infrastructure. It is the runtime governance layer between an agent and the systems it acts on.

01Authorizes actions against policy before they execute.
02Records decisions with the reasoning behind each outcome.
03Produces replayable evidence of every run.
Agent Proposes an action
Intent Structured action proposal
Policy evaluation Rules checked deterministically
Authorization Allow · block · escalate
Execution Runs only when allowed
Evidence Decision and reasoning recorded
Replay Reconstructable run record
Highlighted stages are governed by the control plane
Apache 2.0 Open specification Vendor neutral Reference implementation JSON Schema Deterministic replay Policy driven Extensible
Why it exists

Why Agent Control Plane?

Logging

Traditional logging tells you what happened. It is written after the fact and describes events, not the authority behind them.

Observability

Traditional observability measures how systems behave. It reports on health and performance, not on whether an action was permitted.

Control plane

Agent Control Plane governs what AI agents are permitted to do before execution, and preserves replayable evidence of the decision afterward.

The problem

AI agents are moving faster than enterprise controls.

As agents begin using tools, accessing data, triggering workflows, and making decisions across enterprise systems, organizations need more than policies and dashboards. They need a runtime record of what the agent proposed, what policy decided, why the decision was reached, what was blocked, what authority existed, what sources were relied on, and how the run can be reconstructed later.

01

Unrecorded agent actions

Agents may propose or take actions without a structured pre-action record. Teams cannot distinguish what was proposed from what was executed.

02

Unauthorized tool use

Without a policy gate, agents may invoke tools or access systems beyond their intended authority scope.

03

Missing blocked-action records

When a proposed action is denied, the denial often disappears silently. There is no durable record of what was blocked or why.

04

Opaque policy decisions

Allow, block, and escalate outcomes are difficult to review when there is no trace showing which policy rule matched.

05

Ambiguous source reliance

Agents act on data from tools, APIs, files, databases, user input, and model output. Without reliance records, provenance is unclear.

06

Runs that cannot be replayed

After an incident, teams need a structured record to reconstruct what the agent proposed, what was decided, what was relied on, and what was exported.

Category

What is an Agent Control Plane?

An agent control plane is the layer that authorizes an AI agent's proposed actions against policy before they execute, records the decision and its reasoning, and produces a replayable record of the run. It governs behavior at runtime — distinct from the systems that authenticate identity, observe health, or orchestrate work.

Layer Question it answers When it acts Responsibility boundary
Authentication Who is making the request? Before Establishes identity. Does not decide what that identity may do.
Authorization Is this identity allowed this resource? Before, per request Grants access to resources. Not aware of agent intent or run context.
Observability How is the system behaving? During and after Measures health and performance. Does not gate actions.
Logging What happened? After Describes events after the fact. No pre-execution decision.
Workflow orchestration What runs next? During Sequences steps. Does not evaluate policy or preserve evidence.
Agent Control Plane Is this agent action permitted, and can the run be replayed? Before execution, preserved after Authorizes proposed actions against policy, records decisions with reasoning, and produces replayable evidence.
Design principles

Principles that shape the specification.

01

Deterministic over probabilistic

Identical inputs produce identical decisions and fingerprints.

02

Policy before execution

Actions are evaluated and authorized before they can affect systems.

03

Evidence over assertions

Outcomes are backed by recorded decisions and traces, not claims.

04

Replay by default

Every run yields a reconstructable record, not an optional export.

05

Vendor neutrality

No dependency on a specific model, framework, or hosting provider.

06

Open specifications

Records are governed by public, versioned JSON schemas.

07

Portable implementations

The pattern is reproducible across languages and runtimes.

08

Extensible architecture

Policies, adapters, and exporters are designed to be replaced.

Ecosystem

Where it sits in the AI stack.

Agent Control Plane sits between the agent frameworks that decide what to do and the enterprise systems those actions touch. It does not replace either — it governs the boundary between them.

Models
OpenAIAnthropicGeminiOpen-weight models
Agent frameworks
LangChainCrewAIOpenAI Agents SDKMCP
Agent Control Plane
Authorize · Record · Replay Deterministic policy gate, evaluation traces, run records, and replay bundles.
Enterprise systems
GitHubSAPSalesforceDatabasesREST APIs

Example integrations are illustrative. The control plane is model- and framework-agnostic and defines a boundary rather than a dependency.

Capabilities

What the control plane records and enforces.

Agent Control Plane is not an agent framework and not a model runtime. It sits beside an agent and governs the trail around each proposed action — from the policy decision through to the exported record.

A — Policy gate

Deterministic policy gate

Evaluates proposed agent actions against allowed tools, blocked tools, action type, and active authority scope. Identical inputs produce the same result and deterministic fingerprint.

B — Evaluation traces

Policy evaluation traces

Records the ordered policy-rule path behind each decision, including which rule matched and why the final result was allow, block, or escalate.

C — Blocked actions

Blocked-action records

Creates a structured record when an agent action is denied, rather than letting blocked operations disappear from the audit trail.

D — Authority

Authority records

Captures who or what authorized an action, under what scope, and whether the authorization is time-bound.

E — Reliance

Reliance records

Records which tool, file, database, API, user input, or model output the agent relied on, preserving provenance across the run.

F — Validation

Semantic validation reports

Checks internal consistency across run records and replay bundles, including run identifiers, action references, blocked-action links, trace relationships, and missing-output warnings.

G — Redaction

Redacted exports

Creates privacy-safer copies of run records and replay bundles by redacting payloads, targets, final output, and recorded reasons while preserving structure and fingerprints.

H — Signing

Signed replay bundles

Adds optional HMAC-SHA256 export integrity metadata so replay bundles can be checked for modification after export.

I — Tool adapters

Controlled tool execution

Provides a minimal adapter pattern that executes a tool only after the control plane records the action proposal and the policy gate returns allow.

J — Replay

Replay bundles

Exports a reconstructable run record containing the frame, proposed actions, policy decisions, evaluation traces, authority records, reliance records, blocked actions, and final output.

Replay

Why replay matters.

A replay bundle is a reconstructable record of a complete run. Because the decision path is deterministic, the same inputs reproduce the same outcome — which makes a run inspectable long after it finished. Replay turns a single record into practical leverage across the lifecycle.

Incident investigation

Reconstruct exactly what an agent proposed, what was decided, and what executed.

Compliance

Produce durable evidence of authorization and blocked operations for review.

Debugging

Trace a decision back through the ordered rule path that produced it.

Policy validation

Replay historical runs against a revised policy before rollout.

Model comparison

Compare behavior across models against an identical framed task.

Regression testing

Detect drift when policies, tools, or adapters change over time.

Governance

Give oversight teams a consistent artifact to review and attest to.

Architecture evolution

Validate that changes to the control plane preserve prior behavior.

Reference architecture

Simple by design.

The implementation is intentionally minimal. It demonstrates the control-plane pattern without trying to become a full enterprise governance platform.

Agent
Submits action proposals Receives allow, block, or escalate decisions Executes only when allowed Can use existing orchestration code
Control plane
Frames the task Applies lightweight policy configuration Checks policy and authority Records policy decisions and traces Records blocked actions Records reliance Produces run records
Replay and export layer
Generates replay bundles Produces semantic validation reports Creates redacted exports Signs replay bundles Persists records through a filesystem adapter
Implementation
JSON Schema Pydantic models Filesystem adapter Semantic validation HMAC-SHA256 signing CLI

The public reference implementation uses JSON-schema-governed records and Pydantic models, persisted through a filesystem adapter. It is designed for inspection, adaptation, and extension — not as a turnkey platform.

Open source

Open-source reference implementation.

The repository is public so teams can inspect the pattern, run the examples, adapt the schemas, test deterministic policy gating, validate exported records, redact sensitive fields, sign replay bundles, and experiment with controlled tool execution in their own agent workflows.

It exists to demonstrate an architecture pattern. It is intended to be forked, adapted, extended, and referenced — not simply consumed.

Python package
Pydantic object model
JSON schemas
CLI
Test suite
GitHub Actions workflow
Examples
Apache-2.0 license
View the repository
Who it is for

Built for teams moving from agent demos to governed deployment.

CIOs and CTOs

Need visibility into what AI agents are doing before they touch critical systems. Require structured records of proposed actions, authorized actions, blocked operations, and replayable runs.

CISOs and risk leaders

Need blocked-action records, runtime authorization, validation reports, redacted exports, signed replay bundles, and replayable incident trails.

AI platform teams

Need a lightweight control layer that sits beside existing agent frameworks without replacing them. Require a reference implementation they can adapt and extend.

Governance and compliance teams

Need evidence of what happened, what was allowed, what was blocked, what rule path was followed, what the agent relied on, and how the run can be reconstructed.

Developers building agent workflows

Need a small package that can be added beside existing agent code to record action proposals, gate tool execution, and export replayable run records.

Scope

Intentionally narrow.

This MVP is deliberately scoped. It is not:

× A complete enterprise AI governance platform
× An agent framework
× A model runtime
× A hosted dashboard
× A production key-management system
× A substitute for good policy design
× A guarantee that model outputs are correct
× A full compliance automation system
× A complete security boundary by itself
What it is

A public reference implementation for deterministic policy gating, policy evaluation traces, semantic validation reports, redacted exports, signed replay bundles, controlled tool-adapter execution, and replayable agent-run records.

Project

Maturity and status.

This is an early-stage reference project. The table below reflects current status; some fields are placeholders pending the first tagged release.

Current version
0.1.0 · preview
License
Apache-2.0
Specification status
Draft
Reference implementation
Available · Python
Schemas
JSON Schema · draft
Documentation
Overview · Whitepaper
Roadmap
Published
Conformance
Not yet defined