The best LLM for AutoGen
AutoGen's model is conversation: agents talk to each other — propose, critique, revise — until the group converges. Every exchange is another model call, so a single AutoGen run can rack up far more calls than a one-shot prompt, and a provider rate-limit mid-conversation freezes the whole group. CodeBurst gives AutoGen an OpenAI-compatible endpoint that fails over across providers and repairs empty tool turns, so the conversation reaches a conclusion.
Back-and-forth multiplies your call count
A GroupChat where three agents iterate to consensus can fire dozens of calls in a tight window — exactly the burst pattern that trips a provider's per-minute limit. And if a reasoning agent returns an empty tool-synthesis turn, the turn-taking stalls with no error to act on. The richer your agent conversation, the more it depends on a backend that doesn't fall over.
What CodeBurst adds
| Failure | CodeBurst |
|---|---|
| Provider rate-limits mid-conversation | Reroutes to a healthy provider in the same request; the agent's turn completes. |
| Empty tool-synthesis turn | codeburst-agent retries with a corrected format. |
| A critic agent needs more rigor | Give that agent codeburst-swarm for a multi-model vote. |
Configure the LLM
AutoGen (AG2) treats any OpenAI-compatible endpoint via api_type: "openai":
import os
from autogen import LLMConfig, ConversableAgent
llm_config = LLMConfig({
"model": "codeburst-agent",
"api_type": "openai",
"base_url": "https://codeburst.ai/api/v1",
"api_key": os.environ["CODEBURST_API_KEY"],
})
assistant = ConversableAgent("assistant", llm_config=llm_config)
# give each agent its own LLMConfig if you want different models per role.Your agents and conversation patterns are unchanged — only the endpoint moves, and the whole group gains provider failover and tool-call repair.
Get started
Get an API key How agent failover works →FAQ
How do I set a custom LLM in AutoGen?
An LLMConfig with api_type="openai", base_url="https://codeburst.ai/api/v1", model="codeburst-agent" and your key.
Why does a multi-agent chat stall?
Back-and-forth trips rate limits and empty tool turns; CodeBurst fails over and repairs them.
Different models per agent?
Yes — give each agent its own LLMConfig on the same endpoint.