tea - testing and evaluation for agents

Test and evaluation tool for agentic workflows. Run an agent against a task, assert what it did and measure how it performed.

Getting started

Requirements

• Python 3.11+
• Docker

Install

pip install teaval

To install from source instead:

pip install -e .

Or with uv:

uv sync

Write a test

from tea import (
    Agents,
    check_diff_added,
    check_file_changed,
    check_run_score,
    check_skill_used,
    iterations,
)


@iterations(3, max_failures=1)
def test_add_minus_operation():
    run = (
        Agents.ClaudeCode.prompt("Add 'minus' operation support to the calculator")
        .workdir("examples/demo_calculator")
        .run()
    )

    check_diff_added(run, "calculator.py", "def minus")
    check_diff_added(run, "test_calculator.py", "def test_minus")
    check_file_changed(run, "README.md")
    check_skill_used(run, "add-new-operation")
    check_run_score(
        run,
        (
            "Does this diff correctly implement a subtraction operation "
            "without breaking the existing structure? "
            "Score 1.0 if the implementation is clean and complete, "
            "0.5 if it works but has style issues, "
            "0.0 if it is broken or missing."
        ),
        min_score=0.7,
    )

Run tests

# Run all tests in a folder
tea examples/demo_calculator/agent-tests

# Run one file, repeated 5 times to measure pass rate
tea examples/demo_calculator/agent-tests/test_add_new_operation.py -n 5

# Run a single test function
tea examples/demo_calculator/agent-tests/test_add_new_operation.py::test_add_minus_operation

See examples/demo_calculator/agent-tests for runnable examples.

Metrics baselines

Export per-test metrics (average token count, step count, duration, and success rate) to JSON, then compare later runs against it to catch regressions. --baseline is warn-only — it flags any metric that worsened by more than 10% but never changes the exit code.

tea examples/demo_calculator/agent-tests -n 3 --summary-export baseline.json
tea examples/demo_calculator/agent-tests -n 3 --baseline baseline.json

Checkers

Checkers are the assertions of a tea test: each one inspects an agent run, records a pass or a fail, and returns its result so you can build on it.

Agents

ClaudeCode

Agents.ClaudeCode
    .prompt("Add 'minus' operation support to the calculator")
    .workdir("examples/demo_calculator")
    .run()

Setup ANTHROPIC_API_KEY (API-key auth, recommended). A long-term token (CLAUDE_CODE_OAUTH_TOKEN, generated with claude setup-token) is also supported.

To pick a model, pass its identifier to with_model(); the default is claude-haiku-4-5-20251001:

Agents.ClaudeCode.with_model("claude-opus-4-7").prompt("...").run()

Pi

Pi requires a PiConfig:

from tea.agent import PiConfig

Agents.Pi.with_model(PiConfig(
    model="ollama/qwen2.5-coder:1.5b",  # required
    api_key_env="MY_API_KEY",           # host env var forwarded into container
    base_url="http://host.docker.internal:11434/v1",  # OpenAI-compatible endpoint
))

base_url is used verbatim inside the container, so a provider running on your machine must be addressed as http://host.docker.internal:<port> — not localhost, which resolves to the container itself. Setting base_url makes tea declare host_network_access, which is what makes that name resolve; the address is still yours to write.

Skills

Tea discovers project skills automatically from conventional, in-workdir paths and copies them into the agent's native skill directory in the container.

.agents/skills is the cross-agent convention — put a skill there and every agent picks it up. Missing default paths are silently skipped. If the same skill name appears in multiple sources, the earlier (agent-specific) source wins.

Override the defaults with skills_sources:

# One or more custom source paths (relative to the run's workdir).
Agents(skills_sources=["eval-fixtures/skills"])

# Per-run override.
Agents.ClaudeCode.prompt("...").workdir(".").skills_from(["custom/skills"]).run()

# Explicit opt-out: load no skills at all.
Agents(skills_sources=[])

Each skill is a directory containing a SKILL.md file with frontmatter:

---
name: add-new-operation
description: How to add a new arithmetic operation to the calculator.
---

Step-by-step instructions live in the body of the file.

Development

uv sync --group dev
uv run pre-commit install

Run checks manually:

uv run pre-commit run --all-files
uv run pytest tests/unit -q