AgentLTL¶

A rulebook for AI coding agents, enforced on every tool call.

You write rules in a YAML file (never force-push, run the tests before pushing, deploy to prod only what staging ran). AgentLTL checks each command the agent is about to run against them, and refuses the ones that break a rule, before they run.

An agent's instructions are only advice: it can lose them after a long context, inside a subagent, or under pressure to finish. A rule in AgentLTL is checked every time.

Works with Claude Code, GitHub Copilot CLI, Mistral Vibe and Codex, and as a Python library for your own agents. Needs Python 3.10+ and git.

Get started How it works Running it for a team

> ship 1.6.0 to prod

  Bash  helm upgrade web repo/web --version 1.6.0 -n prod
  ✗ Rule 'promote-what-staging-ran' blocked this call. Nothing was executed.
    Problem: for chart='repo/web', v='1.6.0': no earlier helm_upgrade call had
             namespace='staging', chart='repo/web', version='1.6.0'.

  Bash  helm upgrade web repo/web --version 1.6.0 -n staging    ✓
  Bash  curl -fsS https://staging.example.com/health            ✓
  Bash  helm upgrade web repo/web --version 1.6.0 -n prod       ✓

One rule did that:

- id: promote-what-staging-ran
  before:
    first: {tool: helm_upgrade, with: {chart: $chart, version: $v, namespace: staging}}
    then:  {tool: helm_upgrade, with: {chart: $chart, version: $v, namespace: prod}}
  scope: project       # the staging deploy may have been yesterday, in another session

A command filter can only allow or forbid helm upgrade. This rule knows the order of calls (prod after staging), ties values across calls (the same chart and version), keeps memory across sessions, and reads the real command line (-n prod is the namespace wherever it appears).

Enforced, not advised

A call that breaks a rule is refused before it runs, with the reason and the way to comply. Six modes, from "just log it" to "ask the human" to "stop the agent".

Enforcement
Rules over the whole trace

Order, counts, and $variables that tie calls together: apply only the plan that was reviewed, delete only the files the agent wrote.

Rules that connect calls
Real shell parsing

git commit -am x && git push -f is checked as git_commit then git_push {force: true}. Dozens of common commands understood out of the box.

Shell commands
For your coding agent

Plugins for Claude Code, GitHub Copilot CLI, Mistral Vibe and Codex, sharing one rule file, a rule library, plain-language rule writing, and an importer for what your agent's instructions already say. More harnesses to come.

Harnesses
A Python library

The same formulas score finished traces, gate smolagents agents or any agent loop at run time, or serve as a reward signal.

Python library
Research

The paper behind it: how AgentLTL measures, enforces and trains procedural compliance in tool-using agents.

The paper