agent-in-the-loop

A lightweight Python client for the Agent In The Loop (AITL) confidence evaluation API. Attach a callback to your LangChain / LangGraph run, then call evaluate_confidence() when you want a score. Context, trace ID, and agent name are picked up automatically — no manual wiring.

This library cannot be used without an API key from trellar.io.

Create an account and API key

trellar.io is the only place that issues API keys for this library. Create an account there, then generate an API key from the dashboard. Without that key, evaluate_confidence() cannot authenticate and the client will not work.

Then pass the key into the SDK (see Environment Variables):

• evaluate_confidence(api_key="..."), or
• AGENT_IN_THE_LOOP_API_KEY in the environment

Installation

pip install "agent-in-the-loop[langchain]"

The langchain extra is required because agent runs are captured via a LangChain callback handler (get_agent_guard). Requires Python 3.9+.

Quick Start

from agent_in_the_loop import get_agent_guard, evaluate_confidence

# agent_name must be a stable, unique name for this agent graph — the
# backend uses it to track the graph's network profile across runs.
guard = get_agent_guard("research-agent")

graph.invoke(inputs, config={"callbacks": [guard]})

# context, trace_id, and agent_name are picked up from the guard
result = evaluate_confidence()
print(result.score)        # int, 1-10
print(result.explanation)  # str, human-readable reasoning

Where to call evaluate_confidence

Call it from a graph node (or after invoke()), at the point in the run you want scored. The payload is the events captured so far — later nodes are not included.

There are two ways to use the result:

1. Gate — validate before the graph continues

Put the call on an edge you do not want the graph to cross until AITL has scored the run. Use result.score / result.explanation to decide whether to proceed or stop.

def confidence_gate(state):
    result = evaluate_confidence()
    if result.score < 7:
        return {**state, "halt": True, "reason": result.explanation}
    return {**state, "halt": False}

Wire that node in front of the next step, and only continue when the score is acceptable.

2. Observe — send a validation, do not restrict the graph

Put the call anywhere you want a score recorded (a node, or after invoke()). Store or log result if you want it; do not branch on it. The graph continues either way.

def report_confidence(state):
    result = evaluate_confidence()
    return {**state, "confidence_score": result.score, "confidence_explanation": result.explanation}

Environment Variables

The SDK always talks to the managed AITL backend at https://api.trellar.io — this is fixed and cannot be overridden via an environment variable or function argument.

The API key itself is created only at trellar.io. Once you have it, you can pass it to evaluate_confidence(api_key=...) or set it as an environment variable so you do not pass it on every call:

export AGENT_IN_THE_LOOP_API_KEY=your-api-key

result = evaluate_confidence()  # api_key read from the env var

API Reference

get_agent_guard

get_agent_guard(agent_name: str) -> BaseCallbackHandler

Returns a LangChain callback handler bound to agent_name. Pass it to graph.invoke(..., config={"callbacks": [guard]}).

Raises:

• ValueError — if agent_name is empty or blank

evaluate_confidence

evaluate_confidence(
    *,
    api_key: str | None = None,
    timeout: float = 30.0,
) -> AgentLoopResult

context, trace_id, and agent_name are resolved automatically from the active guard created by get_agent_guard — there is no way to pass them manually. Requests always go to https://api.trellar.io; callers cannot redirect them.

Raises:

• ValueError — if no active guard is found, its trace_id cannot be resolved, or api_key is missing
• requests.HTTPError — on non-2xx HTTP responses

AgentLoopResult

A frozen dataclass with two fields:

License

MIT — see LICENSE for details.