Tasks 0.1.0: the first release

Poisson Tasks makes tool-use tasks for AI agents at a difficulty you set. Every task comes from a seed, has one right end state, and is checked by its own verifier, so no judge model decides what counts as a pass. Python 3.11 to 3.13, on macOS and Linux. It is open source, under the MIT licence.

If you build agents, point Tasks at the tools you already have: an OpenAI tools file, an MCP server or tool list, or an OpenAPI document. It builds tasks on those tools, answers every call from a simulator shaped by your schemas, so your backend is never called, and gives you a card showing where your setup starts to break.

If you train models, the same tasks export to verifiers, checkable at the difficulty you pick.

Start. pip install poisson-tasks, then poisson-tasks with no arguments. It opens a guided session: setup, a preview of one task, a first run and your first card. A first run is 30 episodes and costs about two cents on gpt-5-nano, and the run shows its plan, its expected cost and its cap before it spends anything.

What’s in it. The poisson-tasks command with seven verbs (init, generate, run, card, verify, export, doctor); five providers (OpenAI, Anthropic, Ollama, vLLM and any OpenAI-compatible endpoint); tasks on your own tools, with every call answered by a simulator shaped by your schemas; a run engine that stops at the cap, resumes after Ctrl-C without running anything twice, and lists every file it leaves on your machine; and the two published reference cards, gpt-5-nano and Qwen3-8B at 4,200 episodes each.

The dials. Nine. Depth is measured on both published cards and makes tasks harder for gpt-5-nano. Faults, history and shape shipped from a pre-registered probe on the same two models (60 tasks and 2 attempts per arm, strict pass@1, lowest level against highest). The decoy-tools dial is experimental: it did not make tasks harder for either published model. Similarity, block, arguments and spread are available and retested in v2. Each dial’s figures, and three things about how the probe ran, are in knobs and arms.

An example run. The Petstore API on your own tools with gpt-5-nano, 10 tasks at each level: with no failed calls the strict pass rate was 0.90, and with four it was 0.30. It is one run, not a result about the model.

Known limits. A card from your own run can’t be read beside the published cards yet: the comparison refuses in one line until the next release measures them again. OpenAI shuts down the gpt-5-nano-2025-08-07 snapshot on 2026-12-11. verify doesn’t judge tasks made on your own tools yet; a run’s card judges every one of them by replay. Windows isn’t supported.

Next. The reference cards measured again on current models, a Claude model among them; the four dials that didn’t ship, retested on those models; verify for tasks on your own tools.

The docsThe full release notesThe repositoryPyPI