Break is available: it finds where your agent fails before your users do. Read the launch
Prove is available: it tells you whether a change to your agent changed how often it fails. Read the launch
Poisson Labs

Changelog · Break · 2026-11-16

Break 0.1.0: the first release

0.1.0

Poisson Break finds where a tool-using agent fails. It runs an agent you already have, many times, while it makes the agent’s tool calls fail the way they fail in production, keeps only the failures that hold up against an untouched control, cuts each down to the smallest change that still breaks the agent, and writes it out as a finding a fix can be checked against. Python 3.11 or later, on macOS and Linux. 0.1.0 is an alpha. It is open source, under the MIT licence.

Start. pip install poisson-break, poisson-break init --example desk, poisson-break hunt --yes and poisson-break confirm --yes: three findings on a help desk agent, with no key and no network once it is installed. On your own agent, poisson-break init asks where your agent and its test cases are and, for each tool it finds, whether calling it changes anything, and writes break.toml. Run it against a test environment: Break calls your tools for real.

What’s in it. The poisson-break command with eight verbs (init, doctor, hunt, confirm, findings, replay, recheck, catalog). Ten tool faults: error, timeout, slow, rate_limited, empty, cut_off, malformed, stale, lost_reply and planted_instruction. Five changes to a request’s wording. Two search strategies, screen and random, neither of which learns. A confirm stage on fresh seeds, where every candidate is tested against one shared control with Trace’s exact test and corrected across the candidates. Findings as plain files, each with the command that replays it, and recheck --yes --check as a gate for CI. Five packaged examples, each written into a folder by init --example NAME.

What it does not do. 0.1.0 runs an agent it can call as a Python function that wraps its own tools; it has no adapter for an agent framework. Changes to how a conversation goes and to the agent’s own model calls are named in the catalogue and are not applied in this release. Break grades nothing with a model, computes no statistic of its own (the test is Trace’s and the correction is Prove’s), sends nothing anywhere, and never says that an agent which survived a hunt is safe.

MATH ASINFRASTRUCTURE.

Talk to us