Agent Replay Bundle | Portable replay records for AI-agent runs
Replay/Agent Replay Bundle

Agent Replay Bundle

Portable replay records for AI-agent runs.

A public schema and reference validator for packaging AI-agent runs into portable evidence artifacts. A replay bundle records the task frame, action proposals, policy decisions, evaluation traces, blocked actions, authority records, reliance records, final output, validation report, redaction metadata, and signature metadata.

Apache 2.0 Bundle schema Redaction helper Signing helper CLI Tests
Agent replay bundle signed
run_frame
action_proposals
policy_decisions
evaluation_traces
blocked_actions
reliance_records
validation_report
signature
The problem

Agent traces are fragmented across tools and systems.

Agent runs often span prompts, tools, APIs, databases, files, workflow systems, and policy checks. After an incident or review request, teams may not have a single portable artifact that reconstructs what happened.

01

Logs are scattered

Agent activity lands in prompt logs, tool logs, API logs, and workflow systems, with no single place that reconstructs the run.

02

Tool calls are framework-specific

Each agent framework records tool calls differently, so evidence cannot move cleanly between systems or teams.

03

Blocked actions may be missing

When an action is denied, the denial often never reaches durable storage, leaving gaps in the record.

04

Source reliance is hard to reconstruct

Which file, API, or model output the agent relied on is rarely captured in a form that survives the run.

05

Exports lack integrity metadata

Ad-hoc exports carry no consistent proof of integrity, no redaction record, and no validation result.

06

Runs cannot be reconstructed later

Without one portable artifact, reconstructing an agent run after the fact depends on stitching together systems that may have moved on.

What it does

A portable evidence package for agent runs.

A — Frame

Run frame

The task context the run began from, including inputs and configuration.

B — Proposals

Action proposals

Every action the agent proposed during the run, before any decision.

C — Decisions

Policy decisions

The allow, block, or escalate outcome recorded for each proposed action.

D — Traces

Policy evaluation traces

The ordered rule path behind each decision, showing which rule matched.

E — Blocked

Blocked actions

Structured records of the actions that were denied, and the reasons why.

F — Authority

Authority records

Who or what authorized an action, and under what scope.

G — Reliance

Reliance records

The tools, files, APIs, and outputs the agent relied on across the run.

H — Output

Final output

The result the agent produced once the run completed.

I — Validation

Validation report

The semantic consistency check across the bundle's records.

J — Redaction

Redaction metadata

What was redacted from the exported copy, while preserving structure.

K — Signature

Signature metadata

Optional integrity metadata so the bundle can be checked for modification.

Why it is valuable

One structure for agent-run evidence.

Creates a common structure for agent-run evidence
Supports audit and incident reconstruction
Separates runtime evidence from any one framework
Makes blocked actions and reliance visible
Provides a basis for validation, redaction, signing, and governance review
Can be emitted by Agent Control Plane or adapted by other systems
How it fits in the stack

The Replay layer of the Open Control Stack.

Agent Replay Bundle is the Replay layer. It packages the output of runtime control into a portable artifact that can be validated, redacted, signed, stored, or transformed into a business-facing evidence pack.

Declare Control Replay Evidence
What it is not

A portable artifact — not a platform.

×Not an agent framework
×Not a replay viewer
×Not a model runtime
×Not a hosted audit platform
×Not proof that the model output is correct
×Not a replacement for application logs or security controls
Open-source contents

What is in the repository.

Replay bundle schema
Pydantic object model
Validator
Redaction helper
Signing helper
CLI
Example bundles
Tests
Documentation
View on GitHub
Who it is for

Built for teams that need portable agent-run evidence.

Security and incident teams

Need a single portable artifact that reconstructs an agent run during investigation.

Governance and compliance teams

Need consistent, framework-independent evidence of what an agent run contained.

AI platform teams

Need a common export format their control plane and tools can emit and consume.

Auditors and reviewers

Need to inspect blocked actions, reliance, and decisions without access to live systems.

Developers building agent workflows

Need a schema and validator to package runs into replayable evidence.

Get started

Make agent runs reconstructable.

View on GitHub