Model and Context Management
Up to this point, we’ve mostly relied on the default GPT-4o model when building agents with OpenAI Agents SDK. One of the most powerful features of OpenAI Agents SDK, however, is its model-agnostic design. In other words, you are not locked into a single model provider or configuration. This flexibility becomes especially valuable in complex workflows or multi-agent systems, where different stages of the process might benefit from different capabilities or cost/latency.
Consider a multi-agent system that begins with a triage agent, whose simple job is to ask simple questions to the user. Using GPT-4 for this task may not be appropriate. Instead, we may want a simpler model or a non-OpenAI open sourced model, such as LLaMA, which costs next to nothing. Meanwhile, a more demanding task, such as performing research or math operations, may require advanced Anthropic models. At the same time, a brainstorming agent may need the standard GPT-4o LLM...
