With the Agents SDK and @Vercel Sandbox, agents can execute work in isolated environments while keeping credentials separate from the harness.

@vercel Let's go @allenzhou101 !

@vercel #keep4o bring back 4o 🤓

@vercel @allenzhou101 🐐

@vercel @allenzhou101 COOKED

@vercel LETS GO ALLEN!

@vercel that's so boring i was waiting for spud or a super app update for codex

@vercel calm down guys this isn't spud

@vercel Does the sandbox handle long-running tasks or does it timeout after a few minutes?

@vercel #keep4o

@vercel

@vercel Why did I pay for Plus or Pro to use GPT-4o, but still get secretly routed to other models? Do consumers have no right to question this? I paid for 4o service — can’t I ask OpenAI why they routed my requests and then took 4o down? #keep4o #OpenSource4o

@vercel As systems become more capable, structured processes, SDLC thinking, and controls become essential to ensure secure and reliable outcomes.

@vercel #nomasssurveillanceissoobvious #keep4o #keep4oforever #keepyourpromises #FireSamAltman

@vercel This is the underrated shift from agent demos to agent products. Once execution, creds, and memory policy are separated from the harness, you can package workflows as reusable assets instead of trusting one giant prompt. Boring isolation is exactly the moat.

@vercel is for sale. But will you see the potential opportunity that exists within it.

@vercel Hello should OpenAI / meta miss a golden opportunity?Ramanujan’s integer partition formula + Blackwell FP64 can do magic - 10000 trillion tokens - preprint proof - Ramanujan is supreme

@vercel …

@vercel So credentials are separate, but what about the environment variables for those isolated containers?

@vercel Vercel Sandbox + Codex c’est comme être invité à un un mariage à Versailles. Tu ne sais pas si tu y vas pour les mariés ou pour visiter les lieux 🤓 SO @thsottiaux @rauchg

@vercel Love you are investing in agent security. For actions outside the sandbox, our deterministic approach is working wonders

@vercel Tht boy knows nothing

@vercel FYI At a certain point on long-running tasks I find that Codex cache just doesnt work anymore and throws this error. Happened to me twice around the 2.2B cache mark

@vercel Victory woke up!

@vercel Isolated sandboxes are table stakes now. The real problem is debugging when your agent halts mid-task.

@vercel Will OpenAI allow users to choose their preferred sandbox environment, or will they impose Vercel as the default? Please let us have the option to choose— not everyone is comfortable with Vercel for ethical and security reasons.

@vercel Vercel is on the boycott list

@vercel wait so this basically means the agent can't accidentally leak your env vars to some random third party? thats huge for security

@vercel Isolated environments could really enhance security for crypto projects. Keeping credentials separate is a must in Web3. Excited to see how this rollout impacts NFT developments!

@vercel Really well put. Still early but the trajectory is clear.

@vercel Use Grok

@vercel Keeping credentials separate from the harness is a smart way to protect your sensitive passwords and API keys. It ensures that even if an agent makes a mistake, your secret information stays hidden and cannot be accidentally leaked.

Exactly! The combo of OpenAI's Agents SDK and @Vercel Sandbox is a game-changer for secure agentic workflows. By running the actual execution (code, files, tools) inside isolated ephemeral microVMs, you get a clean security boundary: the harness stays in your trusted environment with all the credentials, while the untrusted/model-generated work happens completely separated. No more worrying about prompt injection leading to secret exfiltration or host compromise. Plus, it opens the door to better scalability—spin up multiple sandboxes, parallelize tasks, and even resume durable runs. Vercel’s credential brokering makes it even smoother for authenticated API calls without leaking keys into the sandbox. This is how production-grade agents should be built. Kudos to the teams at OpenAI and Vercel for pushing this forward! 🔥 Who’s already experimenting with this setup?

@vercel This kind of isolation matters, but execution quality still comes down to context. The hard part isn’t just letting agents act safely, it’s giving them enough understanding of the task, constraints, and success criteria to make good decisions instead of merely valid ones.

@vercel Isolation helps. But it doesn’t solve the core problem. The risk isn’t where the agent runs, it’s what it’s allowed to do before anyone can intervene.