Schwarm - A Free-Form Agent Framework

An Opinionated Agent Framework inspired by OpenAI's swarm.

Incredibly simple yet amazingly deep, it literally wrote parts of itself.

What’s its pull? Literally manipulate everything. No state graphs that need to make sense, no hidden prompts... everything works exactly as you define it. Want to split the agent system in half after 7 rounds, reverse the flow of agents while only communicating in emojis? Yeah. No problem.

Because I’m creative, I called it Schwarm - the German word for swarm.

Features

• Extend your agents with "Providers"

  • Like with the zep_provider which integrates zep (https://github.com/getzep/zep) into an agent, giving it near-infinite memory via knowledge graphs.
  • Use hundreds of LLM providers.
  • Budget tracking.
  • Token tracking.
  • Caching.
  • And so much more.

• Don't waste your time designing your agent state machine in so much detail you end up building a normal static service by accident

  I’ll never understand why people decide on using agents just to remove everything agentic from them with state-graphs so complex the graph itself has started to predict tokens.

  • Let agents be agents.
  • Give them dynamic instructions.
  • Give them dynamic tools/functions.
  • Heck, make everything dynamic. Send every completion call to 50 LLMs and have another Schwarm instance rate them? No problem. Good luck doing that in LangGraph without losing your sanity.

• Extensive logging and visualization

  • Tell your agents to wait for your approval after every step.
  • Log everything in a way that’s actually readable.

• Lightweight, with no memory overhead

  • Agents aren’t memory-hogging objects idling on VRAM.
  • It’s basically one agent cosplaying as many, switching configurations dynamically.
  • (That said, knowing my coding skills, there are probably a few million memory leaks in there.)

• Crazy Use Cases with A Crazy Agent Framework

  • The light-hearted notebooks "The Teachings of Schwarm" (in /lessons) showcase theoretical agent concepts, everything the framework can do, and my random ramblings - all in one. Just hit play. Science definitely went too far.

• Feature parity to other agent frameworks and it's amazing

  • Implement SemanticKernel Agents with Schwarm and watch Schwarm outperforming SemanticKernel Agents, thanks to tool calling strategies, fallback mechanism and all that jazz.

---

Quickstart

1. Install Schwarm with pip:

pip install schwarm

2. or if you clone the repo and want to build it yourself:

 uv venv
 uv sync --all-groups
 poe miau

3. Export your OpenAI API key:

   export OPENAI_API_KEY=sk-xxx

4. Create your agent:

   stephen_king_agent = Agent(name="stephen_king69", configs=[LiteLLMConfig(enable_cache=True), ZepConfig()])

   Mr. Stephen King is ready to rock! And he has his cache with him! All in one line!

   (Caching means that every message interaction will be cached, so if you send the same exact prompt to the LLM, you will receive the cached answer instead of a newly generated one. Saves money and lets you debug!)

5. Instructions, pls!

   Tell it what to do with dynamic instructions that can change every time it’s the agent’s turn again. Carry objects and other data from agent to agent and step to step with the help of context_variables.

   def instruction_stephen_king_agent(context_variables: ContextVariables) -> str:
       """Return the instructions for the user agent."""
       instruction = """
       You are one of the best authors in the world. You are tasked to write your newest story.
       Execute "write_batch" to write something down to paper.
       Execute "remember_things" to remember things you aren’t sure about or to check if something is at odds with previously established facts.
       """
       if "book" in context_variables:
           book = context_variables["book"]
           addendum = f"\n\nYour current story has this many words right now (goal: 10,000): {len(book) / 8}"
           memory = cast(ZepProvider, ProviderManager.get_provider("zep")).get_memory()
           facts = f"\n\n\nRelevant facts about the story so far:\n{memory}"
           instruction += addendum + facts
       return instruction
   
   stephen_king_agent.instructions = instruction_stephen_king_agent

   Did you see the amazing word count algorithm?

6. The toolbox

   Give your agent skills it wouldn’t have otherwise! Also, pass the stick to other agents by setting them in the agent property of the Result object. Just not in this example… Mr. King works alone!

   With this way of doing handoffs, you can implement every state graph you could also build with LangGraph. But this way, you keep your sanity.

   def write_batch(context_variables: ContextVariables, text: str) -> Result:
       """Write down your story."""
       cast(ZepProvider, ProviderManager.get_provider("zep")).add_to_memory(text)
       if "book" not in context_variables:
           context_variables["book"] = ""
       context_variables["book"] += text
       return Result(value=f"{text}", context_variables=context_variables, agent=stephen_king_agent)
   
   def remember_things(context_variables: ContextVariables, what_you_want_to_remember: str) -> Result:
       """If you aren’t sure about something that happened in the story, use this tool to remember it."""
       results = cast(ZepProvider, ProviderManager.get_provider("zep")).search_memory(what_you_want_to_remember)
       result = ""
       for res in results:
           result += f"\n{res.fact}"
       return Result(value=result)
   
   stephen_king_agent.functions = [write_batch, remember_things]

   (Based on the function name, variable names, types, and docstring, a valid OpenAI function spec JSON gets generated. So, this will only work if your model understands those. Support for other tool specs is coming!)

7. Kick off!

   input = """
   Write a story set in the SCP universe. It should follow a group of personnel from the SCP Foundation and the adventures their work provides.
   The story should be around 10,000 words long, and should be a mix of horror and science fiction.
   Start by creating an outline for the story, and then write the first chapter.
   """
   
   response = Schwarm(interaction_mode='stop_and_go').quickstart(stephen_king_agent, input)

   Let your agent system loose! Don’t worry about losing all your money... this quickstart configuration will ask for your approval before performing any money-consuming task.

---

Showcase snippets

Self learning

Want to see what real agent freedom looks like? Try this:

# Create an agent that can evolve its own instructions
def dynamic_instructions(context: ContextVariables) -> str:
    base_instruction = "You are a creative problem solver."
    
    if "learning" in context:
        # The agent can modify its own behavior based on what it learns
        return base_instruction + f"\n\nLessons learned: {context['learning']}"
    
    return base_instruction

evolving_agent = Agent(
    name="learner",
    configs=[LiteLLMConfig(enable_cache=True)],
    instructions=dynamic_instructions,
)

# Let it learn and adapt
def learn_and_improve(context: ContextVariables, insight: str) -> Result:
    context["learning"] = context.get("learning", []) + [insight]
    return Result(
        value=f"I learned: {insight}",
        context_variables=context,
        agent=evolving_agent  # Keep going with new knowledge
    )

evolving_agent.functions = [learn_and_improve]

Dynamic models

Want a system where multiple specialized agents collaborate on a story? Here's how simple it can be:

# Your writing dream team
writer = Agent("novelist", configs=[
    LiteLLMConfig(model="gpt-4"),
    ZepConfig()  # For that infinite memory
])

editor = Agent("editor", configs=[
    LiteLLMConfig(model="gpt-3.5-turbo"),  # Cheaper for edits
    ZepConfig()
])

critic = Agent("critic", configs=[
    LiteLLMConfig(model="claude-3"),  # Different perspective
])

# Let the writer work their magic
def write_chapter(context: ContextVariables, chapter: str) -> Result:
    # save your result
    text = context.get("story", "") + "\n\n" + prompt
    
    # Hand it off to the editor
    return Result(
        value=text,
        context_variables={"story": text},
        agent=editor  # Smooth transition
    )

# Editor cleans it up
def edit_chapter(context: ContextVariables, text: str) -> Result:
    # Make it shine
    edited_text = text + " [Edited for clarity]"
    
    # Let the critic have a look
    return Result(
        value=edited_text,
        context_variables=context,
        agent=critic
    )

Quickies

The Power of True Freedom

# Want to switch models mid-conversation? Sure!
def switch_models(context: ContextVariables) -> Result:
    if context.get("complexity") > 0.8:
        agent.update_config(LiteLLMConfig(model="gpt-4"))
    else:
        agent.update_config(LiteLLMConfig(model="gpt-3.5-turbo"))

Infinite Memory When You Need It

# Remember everything important
zep = agent.get_typed_provider(ZepProvider)
zep.add_to_memory("This is a crucial plot point")
relevant_stuff = zep.search_memory("What happened in chapter 1?")

Budget Control That Makes Sense

# Don't accidentally buy a Tesla in API calls
agent = Agent(configs=[
    LiteLLMConfig(enable_cache=True),  # Save money
    BudgetConfig(
        max_spend=100,
        alert_at=80  # Get warned before it's too late
    )
])

---

Upcoming

• More examples and apps.
• A real documentation.
• Async / true multithreading.
• Integrating dspy.
• An extensive arsenal of providers.
• Dapr backend for true distributed agent shenanigans (also durable functions, aws lambdas etc will be looked at)
• Evaluation/Optimization Framwork. State of the art tools helping your agents a production-ready state. And most of that optimization will be done by an schwarm agent system. Who would have thought!

---

(THIS IS AN ALPHA)
Seriously, this is alpha in every possible sense. It’s a playground for personal agent-based PoCs of all kinds. If you’re looking for an agent framework to perform automatic brain surgeries, you’re in the wrong place.

If you’re looking for something that’ll blow your mind as often as it crashes... hello!

Also, no support during alpha, but I promise the "Teachings of Schwarm" will always work.

Edit: Since Schwarm can already implement parts of itself, the UI is shaping up pretty nicely and quickly, and I estimate beta around q1 2025.