Swarm (experimental, educational)

An educational framework exploring ergonomic, lightweight multi-agent orchestration.

[!WARNING]
Swarm is currently an experimental sample framework intended to explore ergonomic interfaces for multi-agent systems. It is not intended to be used in production, and therefore has no official support. (This also means we will not be reviewing PRs or issues!)

The primary goal of Swarm is to showcase the handoff & routines patterns explored in the Orchestrating Agents: Handoffs & Routines cookbook. It is not meant as a standalone library, and is primarily for educational purposes.

Install

Requires Python 3.10+

Usage

Overview

Swarm focuses on making agent coordination and execution lightweight, highly controllable, and easily testable.

It accomplishes this through two primitive abstractions: Agents and handoffs. An Agent encompasses instructions and tools, and can at any point choose to hand off a conversation to another Agent.

These primitives are powerful enough to express rich dynamics between tools and networks of agents, allowing you to build scalable, real-world solutions while avoiding a steep learning curve.

[!NOTE]
Swarm Agents are not related to Assistants in the Assistants API. They are named similarly for convenience, but are otherwise completely unrelated. Swarm is entirely powered by the Chat Completions API and is hence stateless between calls.

Why Swarm

Swarm explores patterns that are lightweight, scalable, and highly customizable by design. Approaches similar to Swarm are best suited for situations dealing with a large number of independent capabilities and instructions that are difficult to encode into a single prompt.

The Assistants API is a great option for developers looking for fully-hosted threads and built in memory management and retrieval. However, Swarm is an educational resource for developers curious to learn about multi-agent orchestration. Swarm runs (almost) entirely on the client and, much like the Chat Completions API, does not store state between calls.

Examples

Check out /examples for inspiration! Learn more about each one in its README.

basic: Simple examples of fundamentals like setup, function calling, handoffs, and context variables
triage_agent: Simple example of setting up a basic triage step to hand off to the right agent
weather_agent: Simple example of function calling
airline: A multi-agent setup for handling different customer service requests in an airline context.
support_bot: A customer service bot which includes a user interface agent and a help center agent with several tools
personal_shopper: A personal shopping agent that can help with making sales and refunding orders

Documentation

Running Swarm

Start by instantiating a Swarm client (which internally just instantiates an OpenAI client).

`client.run()`

Swarm's run() function is analogous to the chat.completions.create() function in the Chat Completions API – it takes messages and returns messages and saves no state between calls. Importantly, however, it also handles Agent function execution, hand-offs, context variable references, and can take multiple turns before returning to the user.

At its core, Swarm's client.run() implements the following loop:

Get a completion from the current Agent
Execute tool calls and append results
Switch Agent if necessary
Update context variables, if necessary
If no new function calls, return

Arguments

Once client.run() is finished (after potentially multiple calls to agents and tools) it will return a Response containing all the relevant updated state. Specifically, the new messages, the last Agent to be called, and the most up-to-date context_variables. You can pass these values (plus new user messages) in to your next execution of client.run() to continue the interaction where it left off – much like chat.completions.create(). (The run_demo_loop function implements an example of a full execution loop in /swarm/repl/repl.py.)

`Response` Fields

Agents

An Agent simply encapsulates a set of instructions with a set of functions (plus some additional settings below), and has the capability to hand off execution to another Agent.

While it's tempting to personify an Agent as "someone who does X", it can also be used to represent a very specific workflow or step defined by a set of instructions and functions (e.g. a set of steps, a complex retrieval, single step of data transformation, etc). This allows Agents to be composed into a network of "agents", "workflows", and "tasks", all represented by the same primitive.

`Agent` Fields

Instructions

Agent instructions are directly converted into the system prompt of a conversation (as the first message). Only the instructions of the active Agent will be present at any given time (e.g. if there is an Agent handoff, the system prompt will change, but the chat history will not.)

The instructions can either be a regular str, or a function that returns a str. The function can optionally receive a context_variables parameter, which will be populated by the context_variables passed into client.run().

Functions

Swarm Agents can call python functions directly.
Function should usually return a str (values will be attempted to be cast as a str).
If a function returns an Agent, execution will be transfered to that Agent.
If a function defines a context_variables parameter, it will be populated by the context_variables passed into client.run().

If an Agent function call has an error (missing function, wrong argument, error) an error response will be appended to the chat so the Agent can recover gracefully.
If multiple functions are called by the Agent, they will be executed in that order.

Handoffs and Updating Context Variables

An Agent can hand off to another Agent by returning it in a function.

It can also update the context_variables by returning a more complete Result object. This can also contain a value and an agent, in case you want a single function to return a value, update the agent, and update the context variables (or any subset of the three).

[!NOTE]
If an Agent calls multiple functions to hand-off to an Agent, only the last handoff function will be used.

Function Schemas

Swarm automatically converts functions into a JSON Schema that is passed into Chat Completions tools.

Docstrings are turned into the function description.
Parameters without default values are set to required.
Type hints are mapped to the parameter's type (and default to string).
Per-parameter descriptions are not explicitly supported, but should work similarly if just added in the docstring. (In the future docstring argument parsing may be added.)

Streaming

Uses the same events as Chat Completions API streaming. See process_and_print_streaming_response in /swarm/repl/repl.py as an example.

Two new event types have been added:

{"delim":"start"} and {"delim":"start"}, to signal each time an Agent handles a single message (response or function call). This helps identify switches between Agents.
{"response": Response} will return a Response object at the end of a stream with the aggregated (complete) response, for convenience.

Evaluations

Evaluations are crucial to any project, and we encourage developers to bring their own eval suites to test the performance of their swarms. For reference, we have some examples for how to eval swarm in the airline, weather_agent and triage_agent quickstart examples. See the READMEs for more details.

Utils

Use the run_demo_loop to test out your swarm! This will run a REPL on your command line. Supports streaming.

Core Contributors

Ilan Bigio - ibigio
James Hills - jhills20
Shyamal Anadkat - shyamal-anadkat
Charu Jaiswal - charuj
Colin Jarvis - colin-openai
Katia Gil Guzman - katia-openai