OpenAI's GPT-5.6 guide pushes agents to write their own tool-calling code

OpenAI published a builder's guide to GPT-5.6 aimed at developers writing agents, and the recurring theme is doing more work per model call. The headline feature is programmatic tool calling in the Responses API. Instead of the model asking for one tool at a time and passing every result back through the context window, GPT-5.6 can write a small program that coordinates several tools, filters the intermediate data, keeps only what matters, and decides the next step as the work runs. For tool-heavy tasks that means fewer tokens, fewer round trips, and less step-by-step scripting from the developer.

The guide pairs this with a multi-agent mode, now in beta in the Responses API, where one GPT-5.6 instance spins up subagents in parallel and merges their output. OpenAI compares it to the "ultra" mode in its Codex tool and pitches it for tasks that split cleanly into independent parts, where running them at once cuts wall-clock time. Two smaller controls round it out: reasoning continuity, which lets the model reuse reasoning across turns for better multi-turn quality and cache hits through a reasoning.context setting, and explicit prompt caching, where you mark which prompt prefixes to cache. OpenAI says cache writes are billed at 1.25 times the normal input rate while cache reads stay discounted, so the savings only show up when a prefix is reused enough times.

Why it matters

If you build on the OpenAI API, programmatic tool calling changes how you structure an agent: let the model orchestrate tools in code rather than routing every result back through it. Test it against your own token bills before committing, because the caching math only pays off with heavy prefix reuse.