agentmeter

agentmeter instruments multi-agent AI systems to track execution cost, latency, and outcome at the span level - giving you the economic signal your traces are missing.

Multi-agent systems fail silently and spend invisibly. LangSmith and Phoenix show you what happened. agentmeter tells you what it cost, what failed, and how it compares.

Quickstart

from agentmeter import Tracer, SpanKind, configure
def retrieve(q: str) -> list[str]:
    return ["doc-a"]
configure(log_path="./agentmeter_logs/run.jsonl")
tracer = Tracer("my-planner", SpanKind.AGENT)
with tracer.span("my-retriever", SpanKind.RETRIEVER) as s:
    docs = retrieve("change-order status")
    s.set_cost(0.0002)
    s.set_outcome("success")
print(f"Total cost: ${tracer.finish().total_cost_usd:.4f}")

SpanKind Reference

Configuration

from agentmeter import configure

configure()  # silent stdout default
configure(silent=False)  # print JSON events to stdout
configure(log_path="./agentmeter_logs/run.jsonl")  # JSONL file output

Data Model

ExecutionTrace is the immutable top-level record for one execution and includes execution_id, root_agent_id, started_at, ended_at, outcome, and spans, plus computed total_cost_usd and total_latency_ms. Each Span captures span_id, parent_span_id, caller_agent_id, callee_agent_id, capability, timestamps, token counts, cost_usd, outcome, and optional error_message/metadata. AgentIdentity stores persistent identity fields (agent_id, capability, version, tags, created timestamp) for cross-run attribution.

OpenInference Compatibility

agentmeter SpanKinds align with the OpenInference specification used by Arize Phoenix and LangSmith. Traces can be compared against industry benchmarks using the same vocabulary.

Contributing

Contributions are welcome, especially around tracer integrations, emitter backends, and model validation hardening. Changes to SpanKind require discussion before merge because it is part of the public API and affects downstream compatibility.

License

MIT

agentmeter