OpenAI engineer's setup for running agents without writing the loop

Alex @de1lymoonOriginally posted 2026.10.04

Open original on X

openai engineer shows the exact setup he uses to run agents without building the agent loop himself here's how the setup works: 1. pick the model. GPT-6.1 Sol (gpt-6.1-sol) launched at DevDay on Sep 29. OpenAI says it gets close to GPT-6 Astra on agentic coding, computer use and professional work at roughly one-fifth of the price. 2. create a session. one client.beta.agents.sessions.create () call defines the task, model, tools and environment. OpenAI runs the harness, so you don't build the agent loop yourself. 3. choose the runtime. openai_hosted uses the same sandbox infra as Codex and ChatGPT, self_hosted runs on your infra, and none skips files and commands. 4. connect tools. MCP servers, custom functions, web search and computer use all plug into the same session. 5. keep tool use cheap. tool search loads definitions only when needed, while programmatic tool calling can run calls in parallel and return only useful results to context. 6. load skills. point capability_directories at your skills folder and the sandbox starts with your files, packages, skills and plugins. 7. enable subagents. multi_agent: { enabled: true } lets the main agent split work into parallel tasks with separate contexts. the default limit is 6 concurrent agents. 8. attach vaults. pass vault_ids to give the session and its subagents access to configured MCP tools and credentials. 9. handle memory. sessions keep state across turns, older context gets compacted automatically, and the Agents SDK can save lessons into MEMORY.md. 10. observe and steer. stream events, track subagents and send new input mid-turn to redirect the agent. it's that OpenAI now runs the harness for you: context, orchestration, recovery and subagents your job shrinks to choosing the tools and environment.

This post quotes another post: “DevDay 2026 read: building self-verifying agents with Sol and the Agents API”