OpenAI’s prefix guidance: check how your agent keeps history
OpenAI’s 19 November guidance gives commerce app developers a reason to inspect how their agents carry conversation history between requests. Editing an earlier message or tool result can cause newer…
OpenAI’s 19 November guidance gives commerce app developers a reason to inspect how their agents carry conversation history between requests. Editing an earlier message or tool result can cause newer frontier models’ subsequent reasoning items to be dropped, potentially increasing repeated work, token use and latency.
The official guide recommends appending changes rather than rewriting earlier context. It describes configuration_update items for reasoning effort or output-schema changes, and additional_tools for adding tools while preserving existing definitions and their order.
Measure cache behaviour and reasoning separately
Prompt Cache Diagnostics compares requests and reports the first identified difference on a best-effort basis. It cannot prove that reasoning was reused. The response’s input_transformations instead identifies removed reasoning items and reasons such as prefix_mismatch.
OpenAI also distinguishes its supported compaction from home-grown history rewriting: its compaction can retain subsequent reasoning items even when the prefix changes. None of this establishes a fixed saving for an individual app.
A practical check for merchant-facing agents
Our recommendation is to trace one realistic support or catalogue-editing session through your request builder. Look for code that trims old product results, replaces instructions or swaps the tool list after each turn. Record token use, latency, cache hits and task quality before changing the integration.
Reduce oversized catalogue or order responses before sending them the first time, where that still preserves the information needed for the task. Keep tests for changing merchant instructions: preserving history must not mean ignoring a correction or carrying obsolete permissions into a new action.
Use measured results from your own workload to decide whether a change is worthwhile. The documentation is integration guidance, not a new model announcement or a guarantee that every append-only conversation will be cheaper.
More from AI
See all →
