OpenAI's new Agents SDK gives Astra an isolated sandbox workspace
A breakdown of OpenAI's new Agents SDK design: each agent gets its own sandbox, keeps state and files across runs, gates tools with permissions and guardrails, and resumes the same job after a crash.
openai just quietly changed how long-running agents are supposed to work. the new Agents SDK setup gives GPT-6 Astra an actual sandboxed workspace instead of forcing every task through one giant prompt. [this takes ~5 mins to understand… but the architecture is worth stealing] 1. give each agent its own isolated workspace 2. keep state + files between runs 3. gate tool access with permissions + guardrails 4. resume the same job after a crash instead of restarting it [the flow] → task enters the agent desk → Astra reasons + calls tools → sandbox keeps execution isolated → guardrails verify the result → human only steps in before the irreversible action the interesting part is that failures are expected, not treated like edge cases. tool calls can still fail 3–15% of the time, so the whole system is designed around recovery and review. this is how agents stop being demos and start acting like infrastructure. bookmark this before you build your next one.
This post quotes another post: x.com/gippp69/status/2097696163424014406