Multi-provider workflows

Agents can each run on their own LLM provider. A handoff then also switches the model behind the scenes, which lets you send routing to a cheap model, reasoning to a strong one and summarization to a fast one.

openAIProvider := openai.NewProvider(os.Getenv("OPENAI_API_KEY"))
openAIProvider.SetDefaultModel("gpt-4o-mini")

anthropicProvider := anthropic.NewProvider(os.Getenv("ANTHROPIC_API_KEY"))
anthropicProvider.SetDefaultModel("claude-sonnet-4-20250514")

geminiProvider := openai.NewGeminiProvider(os.Getenv("GEMINI_API_KEY"))
geminiProvider.SetDefaultModel("gemini-2.5-flash")

researcher := agent.NewAgent("Researcher", "You research topics in depth.")
researcher.WithModelProvider(anthropicProvider)
researcher.WithModel("claude-sonnet-4-20250514")

summarizer := agent.NewAgent("Summarizer", "You write short summaries.")
summarizer.WithModelProvider(geminiProvider)
summarizer.WithModel("gemini-2.5-flash")

triage := agent.NewAgent("Triage", "You route work to specialists.")
triage.WithModelProvider(openAIProvider)
triage.WithModel("gpt-4o-mini")
triage.WithHandoffs(researcher, summarizer)

r := runner.NewRunner()
r.WithDefaultProvider(openAIProvider) // fallback for agents without their own

result, err := r.RunSync(triage, &runner.RunOptions{
	Input:    "Research X, then summarize it in three bullets.",
	MaxTurns: 10,
})

SetModelProvider is an alias for WithModelProvider, so existing code keeps working.

How a model is resolved

For every turn the runner resolves the model of the agent that is about to run:

  1. If the agent has its own provider (WithModelProvider), that provider resolves the agent’s model name. A run level RunConfig.Model override is not applied to such an agent, so a model name meant for another provider never leaks into it.
  2. Otherwise RunConfig.Model (when set) overrides the agent model, and RunConfig.ModelProvider - or the runner default provider - resolves it.
  3. An agent model that is already a model.Model instance is used as is.
  4. An agent with a provider but no model name uses that provider’s default model.

Because resolution happens per turn, this works the same for Run, RunSync and RunStreaming, including after a handoff.

Bidirectional flows

Multi-provider setups work with bidirectional handoffs too:

delegator := agent.NewAgent("Coordinator", "You coordinate work.")
delegator.WithModelProvider(openAIProvider)
delegator.WithModel("gpt-4o-mini")
delegator.WithBidirectionalHandoffs(researcher, summarizer)
delegator.AsTaskDelegator()

researcher.AsTaskExecutor()
summarizer.AsTaskExecutor()

See handoffs for the details of the delegation protocol and examples/multi_provider_example for a complete program.