summitsdk 0.1.2

Instrumentation layer for multi-agent AI pipelines

summitsdk

Instrument multi-agent AI pipelines and capture every agent-to-agent interaction as a structured trace. Works transparently alongside your existing Anthropic or OpenAI client -- no changes to your agents or pipeline logic required.

Installation

pip install summitsdk

Requires Python 3.8+. OpenAI support is optional:

pip install summitsdk[openai]

Quickstart

Anthropic

from anthropic import Anthropic
from summitsdk.tracer import Tracer
from summitsdk.exporters.local import LocalExporter

exporter = LocalExporter(output_dir="traces/")
tracer = Tracer(exporter=exporter)

client = Anthropic()
wrapped = tracer.wrap_anthropic(
    client,
    from_agent="planner",
    to_agent="worker",
    task_id="task-001",
    round_trip=1
)

response = wrapped.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1024,
    messages=[{"role": "user", "content": "your prompt"}]
)
# trace written to traces/

OpenAI

from openai import OpenAI
from summitsdk.tracer import Tracer
from summitsdk.exporters.local import LocalExporter

exporter = LocalExporter(output_dir="traces/")
tracer = Tracer(exporter=exporter)

client = OpenAI()
wrapped = tracer.wrap_openai(
    client,
    from_agent="planner",
    to_agent="worker",
    task_id="task-001",
    round_trip=1
)

response = wrapped.chat.completions.create(
    model="gpt-4o-mini",
    max_tokens=1024,
    messages=[{"role": "user", "content": "your prompt"}]
)
# trace written to traces/

How it works

wrap_anthropic() and wrap_openai() return a proxy client. Every subsequent API call through the proxy is intercepted after the response returns, a trace is recorded, and the original response is passed back unchanged. Your agents never know the wrapper is there.

Each call to wrap_* accepts four arguments:

Argument	Type	Description
from_agent	str	Name of the sending agent
to_agent	str	Name of the receiving agent
task_id	str	Identifier for the current task
round_trip	int	Interaction number within the task

Trace schema

Each interaction produces a JSON file:

{
  "trace_id": "550e8400-e29b-41d4-a716-446655440000",
  "task_id": "task-001",
  "from_agent": "planner",
  "to_agent": "worker",
  "content": "response text",
  "model": "claude-sonnet-4-20250514",
  "input_tokens": 412,
  "output_tokens": 156,
  "round_trip": 1,
  "framework": "anthropic",
  "timestamp": "2026-06-10T14:07:00Z"
}

Traces are written to the output_dir specified in LocalExporter. Filenames are generated from task_id, agent names, round trip number, and a short UUID fragment to avoid collisions.

Error handling

Trace recording failures log a warning to stderr and never raise an exception or interrupt the underlying API call. If a trace cannot be written, your pipeline continues unaffected.

[summitsdk] Warning: trace recording failed
for planner->worker: PermissionError: ...

Streaming is not supported. Calling messages.stream() through a wrapped client raises NotImplementedError with a clear message.

Compatibility

summitsdk wraps the raw Anthropic and OpenAI Python clients directly. It works with any framework or orchestration layer built on top of these clients including LangGraph, CrewAI, AutoGen, and custom implementations.

Summit

summitsdk is the instrumentation layer for Summit.

Summit ingests the traces produced by this SDK, scores every agent handoff across a set of collaboration dimensions using a reward model, and automatically rewrites agent system prompts to fix the patterns that score lowest. The reward model compounds in accuracy as more pipelines are observed.

License

MIT