Support for custom Agent

[dmvieira]

Support to implement simple agents that don't need LLMs on core.

Describe the feature

The SDK is really great! I just realize that I only can use LLMs for planning agent next step. But sometimes you need simple agents or better control of planning phase.

An example is an agent that decide next step with simple ML classificatio, NLP task, or business rules based on context received.

[rm-openai]

You don't need an agent for this - you can just write code. For example:

first_agent = Agent(...)
result = Runner.run(first_agent, input)

what_to_do_next = run_my_code(result)

if what_to_do_next.foo:
  # do something
elif what_to_do_next.bar:
  # do something else
else:
  # do a third thing

You can also expose functions like what_to_do_next as tools, if useful.

Feel free to reopen if there are followups!

[dmvieira]

Hum... @rm-openai I will try to explain one more time... I think that every code can be implemented as a function and a Agent is not a function. How can I implement a multi-agent system with custom agents and OpenAI agents?

There are a brunch of features that just "a function" doesn't work for custom agents in this SDK:

Handoffs: A powerful feature to coordinate and delegate between multiple agents.
Guardrails: Run input validations and checks in parallel to your agents, breaking early if the checks fail.
Function tools: Turn any Python function into a tool, with automatic schema generation and Pydantic-powered validation.

Without this kind of integration it's not an "SDK to build agentic AI apps" as proposed, it's just an abstraction for OpenAI usage inside an app.

The OpenAI Agents SDK enables you to build agentic AI apps in a lightweight, easy-to-use package with very few abstractions.

Can you please help me? Because my stack has agents that are not LLMs

[dmvieira]

I don't have permission to reopen 😅