OpenAI Skills + Shell for long‑running agents: patterns and pitfalls
OpenAI’s new Skills and Shell tooling make it easier to ship capability‑scoped, long‑running agents for real backend work, but early adopters report reliability gaps you should engineer around. OpenAI’s cookbook shows how to turn discrete capabilities into reusable Skills that your agent invokes via tool calls, enabling least‑privilege execution and clearer observability ([Skills in API](https://developers.openai.com/cookbook/examples/skills_in_api/)); paired with the “tool‑call render” pattern, this turns a chatty bot into a doer with predictable handoffs ([render pattern explainer](https://dev.to/programmingcentral/the-tool-call-render-pattern-turning-your-ai-from-a-chatty-bot-into-a-doer-4cb2)). For workloads that run minutes to hours, OpenAI’s guidance combines Shell, Skills, and compaction to manage state bloat, retry long steps, and keep transcripts affordable and debuggable ([Shell + Skills + Compaction tips](https://developers.openai.com/blog/skills-shell-tips/)). Plan for rough edges reported by developers: an embedding outage returned all‑zero vectors in text‑embedding‑3‑small, some Assistants API file uploads expired immediately, GPT‑5.2 extended‑thinking had very low tokens/sec for some, and Apps SDK toolInvocation status UI required a widget workaround ([embedding outage](https://community.openai.com/t/embedding-model-outage-text-embedding-3-small-api-ev3-model-name-with-all-0-values/1374079#post_10), [files expiring](https://community.openai.com/t/files-instantly-expiring-upon-upload/1366339#post_5), [slow generation](https://community.openai.com/t/gpt-5-2-extended-thinking-webchat-has-unworkably-slow-token-4-tps-generation/1373185?page=3#post_49), [toolInvocation UI bug](https://community.openai.com/t/bug-meta-openai-toolinvocation-invoking-and-meta-openai-toolinvocation-invoked-not-shown-unless-the-tool-registers-a-widget/1374087#post_1)).