tenro 0.2.1

A modern multi-agent AI testing framework

Tenro

Tenro is a modern simulation framework for testing AI agents. Simulate multi-agent workflows and tool usage without burning tokens.

• No API costs — Tests run offline (no LLM calls)
• Deterministic — Simulate responses, errors, and tool results
• Workflow verification — Check tools, edge cases, and agent behaviours

Install

pip install tenro
# or
uv add tenro

Quick Start

Tenro provides a construct pytest fixture that intercepts LLM and tool calls during tests.

# myapp/agent.py
from tenro import link_agent, link_tool

@link_tool
def search(query: str) -> list[str]:
    ...  # calls external API

@link_agent("Assistant", entry_points="run")
class AssistantAgent:
    def run(self, task: str) -> str:
        ...  # agent loop: LLM calls tools, returns final answer

# tests/test_agent.py
from tenro import Provider, ToolCall
from tenro.simulate import llm, tool
from myapp.agent import search, AssistantAgent
from tenro.testing import tenro

@tenro
def test_agent():
    tool.simulate(search, result=["Simulated Doc"])
    llm.simulate(
        Provider.ANTHROPIC,
        responses=[
            ToolCall(search, query="Find docs"),
            "Summary of docs.",
        ],
    )

    result = AssistantAgent().run("Find docs")

    assert result == "Summary of docs."
    tool.verify(search)
    llm.verify_many(Provider.ANTHROPIC, count=2)

No mocks to configure, no expensive API calls, no flaky tests.

Before / After

Without Tenro — manual mocks, helper functions, boilerplate:

# test_helpers.py - you write and maintain this
def mock_llm_response(content=None, tool_call=None):
    if tool_call:
        message = ChatCompletionMessage(
            role="assistant", content=None,
            tool_calls=[ChatCompletionMessageToolCall(
                id="call_abc", type="function",
                function=Function(name=tool_call["name"], arguments=json.dumps(tool_call["args"]))
            )]
        )
    else:
        message = ChatCompletionMessage(role="assistant", content=content, tool_calls=None)
    return ChatCompletion(
        id="chatcmpl-123", created=0, model="gpt-5", object="chat.completion",
        choices=[Choice(index=0, finish_reason="stop", message=message)]
    )

# test_agent.py
@patch("myapp.tools.get_weather")
@patch("openai.chat.completions.create")
def test_agent(mock_llm, mock_weather):
    mock_weather.return_value = {"temp": 72, "condition": "sunny"}
    mock_llm.side_effect = [
        mock_llm_response(tool_call={"name": "get_weather", "args": {"city": "Paris"}}),
        mock_llm_response(content="It's 72°F and sunny in Paris."),
    ]
    result = my_agent.run("Weather in Paris?")
    assert result == "It's 72°F and sunny in Paris."
    mock_weather.assert_called_once_with(city="Paris")

With Tenro:

from tenro import Provider, ToolCall
from tenro.simulate import llm, tool
from myapp.agent import get_weather, WeatherAgent
from tenro.testing import tenro

@tenro
def test_agent():
    tool.simulate(get_weather, result={"temp": 72, "condition": "sunny"})
    llm.simulate(
        Provider.OPENAI,
        responses=[
            ToolCall(get_weather, city="Paris"),
            "It's 72°F and sunny in Paris.",
        ],
    )

    result = WeatherAgent().run("Weather in Paris?")

    tool.verify(get_weather)
    llm.verify_many(Provider.OPENAI, count=2)
    assert result == "It's 72°F and sunny in Paris."

No patch decorators. No response builders. Just simulate and verify.

How It Works

Tenro's Construct is a simulation environment for your AI agents. Link your functions with decorators, then test with full control:

from tenro import link_agent, link_tool

@link_tool
def search(query: str) -> list[str]:
    ...  # calls external API

@link_agent("Manager", entry_points="run")
class ManagerAgent:
    def run(self, task: str) -> str:
        ...  # LLM calls search tool, summarizes results

During tests, Construct intercepts linked LLM and tool calls and returns your simulated results instead of calling the real provider.

Simulation API

from tenro import Provider
from tenro.simulate import llm, tool
from tenro import ToolCall
from myapp.agent import search, MyAgent
from tenro.testing import tenro

@tenro
def test_verification():
    # Setup
    tool.simulate(search, result=["doc1", "doc2"])
    llm.simulate(
        Provider.ANTHROPIC,
        responses=[
            ToolCall(search, query="docs"),
            "Summary",
        ],
    )

    # Run
    MyAgent().run("query")

    # Verify
    tool.verify(search)                              # at least once
    tool.verify_many(search, count=1)                # exactly once
    llm.verify_many(Provider.ANTHROPIC, count=2)     # exactly twice

    # Access call data
    assert llm.calls()[1].response == "Summary"

Trace Output

Enable trace visualization to debug agent execution:

Set TENRO_TRACE=true in your .env or run TENRO_TRACE=true pytest

🤖 SupportAgent
   ├─ → user: "My order #12345 hasn't arrived"
   │
   ├─ 🧠 claude-sonnet-4-5
   │     ├─ → prompt: "Help customer: My order #12345 hasn't arrived"
   │     └─ ← tool_call: lookup_order(order_id='12345')
   │
   ├─ 🔧 lookup_order
   │     ├─ → order_id='12345'
   │     └─ ← {'status': 'shipped', 'eta': '2025-01-02'}
   │
   ├─ 🧠 claude-sonnet-4-5
   │     ├─ → prompt: "Tool result: {'status': 'shipped', ...}"
   │     └─ ← "Your order has shipped and will arrive by Jan 2nd!"
   │
   └─ ← "Your order has shipped and will arrive by Jan 2nd!"

────────────────────────────────────────────────────────────────
Summary: 1 agent | 2 LLM calls | 1 tool call | Total: 1.24s

LLM Provider Support

Provider	Status
OpenAI	✅
Anthropic	✅
Gemini	✅
Custom	Experimental

Compatibility

• Python 3.11+
• pytest 7.0+

Contributing

Thanks for your interest in contributing!

Tenro is still in the early stages, focused on stabilizing the core API. Pull requests are not being accepted at this time.

You can still help by:

• ⭐ Star the repo to follow progress and help others discover it
• Report bugs (include repro steps + logs if possible)
• Request features (share the use case and expected behavior)
• Ask questions (usage, roadmap, design decisions)

Please use GitHub Issues for discussions and reports.

License

Apache 2.0

Support

• Issues: GitHub Issues
• Email: [e-mail]