GPT-5.6 agent controls and Cursor Builds cut repeat work

OpenAI's GPT-5.6 builder guide puts three agent controls into one architecture: persist reasoning across turns, compact long-running conversations, and use programmatic tool calling for deterministic filtering and aggregation outside the model context. OpenAI reports that Rogo matched its quality rubric while using 21% fewer input tokens with programmatic tool calling. 1

Cursor's Cloud Agent Builds prepare repositories, dependencies, and setup scripts before an agent starts. Cursor says its own environments now boot 10x faster and reach the first token 3x faster. Builds keep the last successful environment active after a failed update, while the dashboard records logs, commit SHAs, and the build used by each run. 2

Try one controlled benchmark on the same repository and task. Record environment startup time, input tokens, and whether the result passes your tests. Put deterministic data plumbing in code, move install-time dependencies into the build, and keep credentials in team or environment secrets.