OpenAI Upgrades Responses API With Persistent Memory and Managed Containers
The Development
OpenAI rolled out three major updates to its Responses API on February 10: Server-side Compaction, Hosted Shell Containers, and Skills standard implementation. Server-side Compaction allows agents to run for hours or days by compressing conversation history instead of truncating it—Triple Whale's agent Moby successfully handled 5 million tokens and 150 tool calls without accuracy loss. The new Shell Tool provides managed Debian 12 environments with persistent storage, networking capabilities, and pre-loaded execution environments including Python 3.11 and Node.js 22. OpenAI also adopted the open agentskills.io standard, using SKILL.md manifests that are portable across platforms. Enterprise AI search startup Glean reported tool accuracy jumping from 73% to 85% using OpenAI's Skills framework. The updates address the core limitation of AI agents losing context during long-running tasks.
Our Take
OpenAI just shifted from selling models to selling infrastructure. By bundling memory management, compute environments, and standardized skills into one API, they're positioning themselves as the operating system for AI agents rather than just another LLM provider. The real insight here isn't the technical capabilities—it's the strategic pivot. While Anthropic focuses on partner integrations and pre-built playbooks, OpenAI is betting on developer velocity and managed infrastructure. This creates a classic build-versus-buy decision for enterprises: OpenAI's turnkey stack versus custom orchestration. The adoption of the open Skills standard is particularly shrewd—it gives developers portability while locking them into OpenAI's high-performance execution layer. This isn't just about better agents; it's about who controls the runtime environment where AI work actually happens.
The Exploit
Action Item
If you're currently building custom state management or sandboxing for AI agents, migrate to OpenAI's Responses API before Q2 2026 budget planning—the infrastructure overhead savings will fund additional agent development projects.