agentpanel

A small, domain-agnostic core for multi-agent LLM workflows on LangGraph:

• Specialists produce artifacts in parallel.
• Reviewers judge each other's artifacts; rejections send the author back to revise, with the feedback, for a bounded number of rounds.
• A coordinator assembles the result, including anything still unresolved.

It runs on local models by default (Ollama), with bring-your-own-key OpenAI and Anthropic support, and gets reliable structured output from small models.

pip install agentpanel            # Ollama only
pip install "agentpanel[cloud]"   # plus OpenAI and Anthropic

Example

import asyncio
from agentpanel import AgentContext, LLMReviewer, LLMSpecialist, Workflow
from agentpanel.config import LLMSettings
from agentpanel.llm import create_chat_model

model = create_chat_model(LLMSettings())  # qwen3:8b on local Ollama by default


class Join:
    async def assemble(self, ctx: AgentContext) -> str:
        return ctx.artifacts["writer"].content


workflow = Workflow(
    specialists=[
        LLMSpecialist(
            name="writer",
            kind="product description",
            model=model,
            instructions="You write concise product descriptions.",
            render=lambda ctx: f"Describe {ctx.inputs['product']} in two sentences.",
        )
    ],
    reviewers=[LLMReviewer("editor", model, "You edit marketing copy for accuracy.")],
    coordinator=Join(),
    max_revisions=1,
)
result = asyncio.run(workflow.run({"product": "a solar-powered kettle"}))
print(result.result, result.unresolved)

Pieces

Configuration

LLMSettings reads AGENTPANEL_* environment variables (subclass it to use your own prefix): PROVIDER (ollama, openai, anthropic), MODEL, BASE_URL, TIMEOUT, REASONING, CONTEXT_WINDOW, AGENT_MODELS, AGENT_REASONING. Cloud keys use the standard OPENAI_API_KEY / ANTHROPIC_API_KEY.

Built for and used by API Shakedown, where the source lives. Issues and contributions: see that repository.

License

MIT