Agent Ledger

Local-first / open source / Codex + Claude Code

Agent Ledger

Make your coding agent check its own work.

Agent Ledger gives coding agents a bounded loop for scope, execution, verification, finding review, and safe correction. You get a more disciplined run and a clear record of what happened.

Run the 15-minute betaOpen the live demo
npx skills add sprintagency-it/agent-ledger --skill agent-ledger --yes --copy

REAL OUTPUT

The human report is evidence of the loop.

The agent reads the same findings before handoff, classifies them, corrects eligible issues, and renders the final state again.

Agent Ledger replay showing run context, status, risk counts, and review evidence
Offline replay from a generated Agent Ledger run. No hosted account required.

01 / WORKFLOW

One task. One bounded correction loop.

Agent Ledger changes how the agent executes the task, not just how the result is formatted afterward.

01

Scope

Define outcome, allowed paths, checks, and boundaries before editing.

02

Execute

Perform the task while recording practical command and file evidence.

03

Verify

Run tests, lint, smoke checks, or the closest available proof.

04

Review

Render machine-readable findings and a PASS, WARN, or BLOCK status.

05

Correct

Classify findings and fix safe true positives inside declared scope.

06

Explain

Hand off what changed, what passed, and what still needs a human.

02 / DECISION ROUTING

Not every signal is a blocker.

Deterministic findings route attention. The agent still has to classify context before it fixes or dismisses anything.

PASS

Normal review

No critical or high signals. Verification and normal code review still apply.

View PASS replay
WARN

Inspect first

A high signal needs evidence-based classification before trust or merge.

View WARN replay
BLOCK

Human decision

A critical signal crosses a boundary that should not be fixed blindly.

View BLOCK replay

03 / START

Install once. Use it with Codex or Claude Code.

Install from the skills directory in your repository root. The skill includes its runtime and prepares the ignored .agent-ledger/ workspace on first use.

  1. Install the skillCopies the workflow and bundled local runtime.
  2. Choose your agentUse the same local workflow in Codex or Claude Code.
  3. Invoke the skillGive it a concrete coding task and expected checks.
quickstart
# 1. Install in the current repository
npx skills add sprintagency-it/agent-ledger --skill agent-ledger --yes --copy

# 2. In Codex
Use $agent-ledger to fix the signup validation bug.

# Or in Claude Code
/agent-ledger Fix the signup validation bug.

# 3. Receive
PASS / WARN / BLOCK
executive-summary.md
review.json
fix-brief.md
replay.html

04 / TWO READERS

Useful to the next agent. Legible to the human.

For the agent

review.json and fix-brief.md turn the end of a task into structured input for classification, correction, or supervisor-agent aggregation.

For the human

executive-summary.md and replay.html show scope, evidence, findings, and unresolved decisions without reopening the original chat.

PUBLIC BETA / V0.3.3

Try it on one real agent task.

Agent Ledger is free, local-first, and intentionally narrow. The beta asks one question: does the loop catch or clarify something useful before handoff?

Start the 15-minute testRead the docs