entropy-loop-core 0.6.0

A Failure Compiler for AI agents: turn failed outputs into lessons, evaluation summaries, and replayable regression cases.

Entropy Loop Core

A Failure Compiler for AI agents. Turn failed agent outputs into regression cases and replay them before the same bug ships again.

Get started · Example · CLI · Architecture · Releases · Contributing

Star the repo if you want to follow the Failure Compiler roadmap.

Get started

pip install entropy-loop-core
entropy-loop replay-demo

Works on Windows, macOS, and Linux with Python 3.10+.

Development setup

Use a virtual environment when working on the repository.

macOS / Linux

git clone https://github.com/koreaelonmusk/entropy-loop-core.git
cd entropy-loop-core
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e ".[dev]"
pytest

Windows PowerShell

git clone https://github.com/koreaelonmusk/entropy-loop-core.git
cd entropy-loop-core
py -m venv .venv
.\.venv\Scripts\Activate.ps1
py -m pip install --upgrade pip
py -m pip install -e ".[dev]"
pytest

Why

AI agents often fail the same way twice.

Entropy Loop Core makes failures reusable: capture the failed output, classify it, compile a lesson, generate a regression case, and replay it — before the same bug ships again.

Task
→ AgentOutput
→ VerificationResult
→ FailureTrace
→ Lesson
→ RegressionCase
→ RegressionSuite
→ Replay
→ Report

The core is deterministic: no LLM calls, no network calls, no hidden state.

Example

Turn a failure into a regression case, then replay it against a fixed agent:

from entropy_loop_core import (
    AgentOutput,
    FailureTrace,
    RegressionRunner,
    RegressionSuite,
    RetryContext,
    Task,
    VerificationPolicy,
    Verifier,
    generate_regression_case,
)

# A verifier built from a policy: non-empty output that contains "status".
verifier = Verifier.from_policy(
    VerificationPolicy(require_non_empty=True, required_terms=["status"])
)

# A past failure (the agent omitted "status") becomes a regression case.
task = Task(id="job-1", instruction="report the job status")
bad = AgentOutput(content="done")
case = generate_regression_case(
    FailureTrace(
        task=task,
        output=bad,
        verification_result=verifier.verify(bad),
        attempt=1,
    )
)


# Replay the case against a corrected agent.
def fixed_agent(task: Task, ctx: RetryContext) -> AgentOutput:
    return AgentOutput(content="status: ok")


report = RegressionRunner().run_suite(
    RegressionSuite(name="job", cases=[case]), fixed_agent, verifier
)
print(report.passed, report.total_cases, report.success_rate)  # 1 1 100.0

Full worked example: examples/json_agent_guard.py.

CLI

entropy-loop replay-demo   # generate a regression case, then replay it as a suite
entropy-loop memory-demo   # compact repeated failure lessons with a MemoryPolicy
entropy-loop pack-demo     # build, save, load, and run a regression pack
entropy-loop agent-demo    # refresh a pack from an agent, then run it
entropy-loop demo          # run the loop: verify → trace → learn → retry → regress
entropy-loop doctor        # health-check the install

memory-demo shows how repeated failure lessons can be compacted with a deterministic MemoryPolicy — see docs/memory-policy.md.

Run a regression pack in CI

Turn captured failures into a portable pack and gate your build on it:

entropy-loop run-pack examples/json_agent_guard.pack.json

run-pack exits non-zero when a known agent regression reappears (0 = pass, 1 = failure, 2 = bad input), making replayable failure checks usable in CI. To gate on your agent's current output, refresh the pack from an explicit local command first (no shell, no secret injection):

entropy-loop refresh-pack input.pack.json output.pack.json -- python my_agent.py
entropy-loop run-pack output.pack.json

See docs/regression-packs.md, docs/agent-adapters.md, and docs/github-actions.md.

What it is / what it is not

It is

• a deterministic failure compiler,
• a structured failure-trace layer,
• a regression replay primitive,
• a small AI-agent reliability tool.

It is not

• a full agent framework,
• model training,
• model-as-judge by default,
• a correctness guarantee,
• a cloud platform.

Architecture

• Verifier applies ordered, deterministic rules and classifies failures.
• EntropyLoop runs an agent, verifies, traces the failure, compiles a lesson, and retries.
• LessonGenerator turns a FailureTrace into a reusable Lesson.
• generate_regression_case pins a failure as a repeatable check.
• RegressionRunner replays a RegressionSuite and returns a report.

Deeper reading: architecture · reliability model · research influences · recording the demo.

Public / private boundary

Open-source the primitive, not the private advantage.

This repository contains only generic reliability primitives — no business logic, proprietary prompts, customer data, secrets, external AI API calls, or network calls. See docs/public-private-boundary.md.

Releases

• v0.6.0 — agent adapter / live pack refresh (current)
• v0.5.0 — regression packs / CI gate
• v0.4.0 — memory policy / lesson compaction
• v0.3.1 — packaging readiness
• v0.3.0 — replay
• v0.2.0 — classification + evaluation
• v0.1.0 — the first Failure Compiler loop

Details in CHANGELOG.md.

Roadmap

• Next (directional) — persistence adapters, richer reports, and broader failure-memory recall.

Full plan in docs/roadmap.md.

Contributing

Contributions are welcome. Keep the core small, readable, and deterministic, and respect the public/private boundary. See CONTRIBUTING.md and CODE_OF_CONDUCT.md.

ruff check .    # lint
ruff format .   # format
pytest          # tests

License

Released under the Apache-2.0 license.