Shell + Skills + Compaction: Tips for long-running agents that do real work

The text explores the transition from single-turn assistants to long-running agents capable of handling complex tasks, facilitated by new agentic primitives developed by OpenAI. These primitives include Skills, which are reusable and versioned instructions that enhance task execution reliability, an upgraded shell tool for executing tasks in controlled environments, and server-side compaction to manage long agentic runs without hitting context limits. Skills act as procedural playbooks for the model to consult, while the shell tool provides a real terminal environment for executing scripts and generating outputs. Compaction ensures continuity in long workflows by managing context windows. The text provides practical tips for utilizing these primitives, such as writing skill descriptions effectively, including negative examples to improve skill routing, and leveraging templates within skills to avoid token inflation. It also emphasizes the importance of designing for long runs, ensuring deterministic skill use, and maintaining strong security postures through network allowlists and domain secrets. The document concludes with build patterns that demonstrate how to combine these primitives to develop effective applications, underscoring the value of Skills, hosted shell, and compaction in creating robust, repeatable workflows.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.