agentlasso

Send your AI agent's production traces to AgentLasso: it learns what your agent actually does, shows what you're not testing, and turns real behavior into golden tests that run in CI.

No dependencies: standard library only. Python 3.9+.

pip install agentlasso

Send a trace

Set your project's API key (AgentLasso → Settings):

export AGENTLASSO_API_KEY="<your-project-api-key>"

Then, wherever your agent finishes handling a request:

import agentlasso

agentlasso.send_trace(
    user_input="I want my money back for order 44182",
    agent_output="Refund of $89.00 issued to your card, 3-5 business days.",
    tool_calls=[
        {"name": "orders.lookup", "arguments": {"order_id": "44182"}, "latency_ms": 210},
        {"name": "stripe.refund", "arguments": {"charge_id": "ch_1P9", "amount": 8900}},
    ],
    session_id="conv_8fa21",
    model="claude-sonnet-5-5",
    background=True,  # don't make the user wait
)

One call per request your agent handled. Send tool_calls in the order the agent called them: trajectory checks, risk, money at stake and coverage are all built on them. arguments can be a dict or a JSON string (as OpenAI-style tool calls carry them).

Background sending

With background=True, send_trace returns at once and a failure is logged to the agentlasso logger instead of raised, so a trace can never break your agent's reply. A normal Python exit waits for pending sends; in a short-lived process (a script, a serverless function), call agentlasso.flush() before returning.

Without it, send_trace waits and returns AgentLasso's response ({"accepted": 1, "traceIds": [...], ...}), or raises agentlasso.AgentLassoError (with .status, the HTTP status code).

Configuration

from agentlasso import Client

client = Client(
    api_key="...",                                   # default: AGENTLASSO_API_KEY
    base_url="https://agentlasso.example.com/api/v1",  # self-hosted; default: AGENTLASSO_BASE_URL, then agentlasso.dev
    timeout=10.0,
)
client.send_trace(user_input="...", agent_output="...")

Already running OpenTelemetry? You don't need this package: point your OTLP/HTTP exporter at the same endpoint. See Send traces.

Links

• Developer guide
• API reference
• Source

Apache-2.0