Responses API: agent tools upgraded
OpenAI pushed major Responses API upgrades for autonomous agents — shell tool support, built-in agent loops, hosted workspaces, context compaction, and reusable skills to streamline agent development. Those additions are squarely aimed at teams building autonomous workflows and production-grade agents. (x.com)
OpenAI published the technical write-up “From model to agent: Equipping the Responses API with a computer environment” on March 11, 2026, crediting engineers Bo Xu, Danny Zhang, and Rohit Arunachalam as authors. (openai.com) The post describes a tight propose-act-observe execution loop in which the model suggests shell commands and the platform executes them in an isolated workspace, returning command outputs back into the next inference step. (openai.com) Hosted shell runtimes are managed containers based on Debian 12 with a default working directory at /mnt/data, no sudo privileges, no interactive TTY support, and preinstalled runtimes including Python 3.11, Node.js 22.16, Java 17, PHP 8.2, Ruby 3.1, and Go 1.23. (developers.openai.com) OpenAI’s Skills are versioned, mountable bundles anchored by a required SKILL.md manifest; developers can upload skills as a directory multipart upload or as a single top-level ZIP and reference explicit skill versions via skill_reference in hosted containers. (developers.openai.com) Server-side compaction runs automatically when a conversation crosses a configured token threshold, emits a compaction output item inline to preserve essential state, and can also be invoked explicitly via the /responses/compact endpoint to produce opaque, compacted conversation items. (developers.openai.com) The Responses API supports creating reusable container objects (example parameters include memory_limit and expires_after anchored to last_active_at) so agents can retain a long-lived execution environment across multiple calls. (developers.openai.com) OpenAI’s developer guidance and examples — published Feb. 11, 2026 — highlight practical patterns and cite early production usage (including Glean as an early skills customer) for combining skills, hosted shells, and compaction to run multi-step, long‑horizon workflows. (developers.openai.com)
