OpenAI DevDay 发布 Agent Stack 全套组件

Morlex @0xMorlex原推发布于 2026.10.02

在 X 上打开原推

OpenAI Agent Stack: stop building chatbots, start building workers OpenAI didn't just ship another model at DevDay it shipped most of the pieces of an agent OS: GPT-6.1 Sol, Agents API sessions, tools, computer use, memory and subagents connect them and you get something much closer to a digital worker than a chatbot Then build around a few hard rules: → vague prompt: give it a job, not "you are a helpful assistant" → agent fails: add the missing tool, not 500 more words of prompt → no API: give it a computer → long history: store state, don't dump it into context → tool returned: verify the result, don't assume it worked the main win is not a smarter model it is a better system around the model The full stack: - Sol handles the reasoning - Agents API keeps sessions alive and recovers from failures - tools turn reasoning into actions - computer use works where no API exists - memory stops the agent from re-reasoning the same things - subagents split independent work in parallel - verification stops autonomy from becoming faster failure the full guide in 10 steps, with code ↓ 1. start with a job, not a prompt 2. give the agent three connections 3. put GPT-6.1 Sol at the reasoning layer 4. move the loop into the Agents API 5. add tools before adding more prompt 6. give the agent a computer 7. build memory as state, not one giant prompt 8. split independent work across subagents 9. add verification and recovery 10. assemble the full stack