AI Sandboxes
Synopsis
AI sandboxes are controlled execution environments where agents can run code, browse pages, inspect files, call tools, render interfaces, or manipulate artifacts without giving the model unrestricted access to the host system. In the World’s Fair material, the idea shows up across agent builders, browser agents, MCP apps, eval runners, cloud deployment workflows, and production assistants: the sandbox is the boundary that lets an agent act while keeping filesystem, network, credentials, browser state, and process execution under policy.
Origin And Context
The pattern comes from operating-system isolation, browser sandboxes, CI runners, notebooks, container platforms, secure code-execution services, and cloud GPU runtimes. The connected sessions make it more agent-specific: Nearform’s agent-building slide deck frames agents as systems that create and operate artifacts; Cloudflare’s Eval++ talk treats evaluation as a repeatable compute primitive; Google’s Chrome DevTools and WebMCP talks put browser/tool surfaces in front of agents; Amazon Nova Act and MCP show structured tool execution; Modal and RunPod sessions point to remote execution environments where APIs, endpoints, and GPUs become the sandboxed substrate.
Why It Matters
Agents need to experiment, test, browse, and inspect state, but the conference evidence repeatedly shows that those actions are only useful when they are observable and bounded. Browser-agent talks emphasize that agents can misread pages or overclaim web search; production talks from OpenGov and Databricks point toward enterprise controls, auditability, and deployment discipline; logging and continual-learning sessions show that failures need to be captured as durable evidence. A sandbox turns agent action into something reviewable: commands, diffs, traces, screenshots, tool calls, resource use, and failure cases can be inspected instead of trusted on faith.
How To Use It
Choose isolation based on what the agent can touch. A low-risk assistant may only need a temporary workspace and subprocess limits; generated code, dependency installation, or untrusted scripts should run in containers, remote runners, or microVM-like environments; browser agents need constrained profiles, explicit download/upload rules, and page-state capture; MCP or ChatGPT app surfaces need iframe and tool-boundary controls; production agents need policy-controlled network access, scoped credentials, quotas, and human-review checkpoints. Capture logs, diffs, artifacts, browser traces, screenshots, evaluation results, and resource usage so the sandbox produces evidence as well as containment.
Where It Is Useful
They are useful in coding assistants, browser and computer-use agents, MCP apps, app builders, data-analysis workspaces, test runners, eval harnesses, educational tools, cloud endpoint deployment flows, and software-factory pipelines. The connected resources stretch the pattern from local IDE workflows and Chrome DevTools-style interfaces to GPU-backed endpoint deployment and enterprise agent operations.
When To Use It
Use a sandbox whenever an agent can execute code, inspect user files, download dependencies, browse unknown sites, call external tools, transform documents, or run scripts generated from model output. Tighten limits when credentials, production data, customer workflows, or browser sessions are involved; loosen them only after the task model, threat model, logging requirements, and rollback path are well understood.
Active Use Cases
- Running generated code and tests before suggesting a patch.
- Browser or computer-use automation with constrained state.
- Temporary workspaces for data analysis and document transformation.
- Reproducible agent task environments for evaluations.
Related Slide Decks
- youtube aHhB3sjGjkI slides — Agents Building Agents - Alfonso Graziano, Nearform (24 extracted slide frames)
Related Scheduled Sessions
- 2026 06 30 pierluca d oro computer use at the edge of the statistical precipice — Computer Use at the Edge of the Statistical Precipice; Pierluca D'Oro (Day 3 — Session Day 2 · 11:10am-11:30am · Computer Use; official schedule)
- 2026 06 30 robert brennan sandboxes aren t optional runtime isolation patterns for coding agents at scale — Sandboxes Aren't Optional: Runtime Isolation Patterns for Coding Agents at Scale; Robert Brennan (Day 3 — Session Day 2 · 3:20pm-3:40pm · Sandbox & Platform Engineering; official schedule)
- 2026 06 30 samuel colvin your agent needs a sandbox not a desert — Your agent needs a sandbox, not a desert; Samuel Colvin (Day 3 — Session Day 2 · 12:05pm-12:25pm · Sandbox & Platform Engineering; official schedule)
- 2026 06 29 tushar jain unlock agent autonomy the runtime for ai native systems — Unlock Agent Autonomy: The Runtime for AI-Native Systems; Tushar Jain (Day 2 — Session Day 1 · 3:45pm-4:05pm · AI Architects: Show my Workflow; official schedule)
- 2026 06 30 abhishek bhardwaj from fork to fleet designing an agent sandbox cloud pt 1 — From fork() to Fleet: Designing an Agent Sandbox Cloud Pt 1; Abhishek Bhardwaj (Day 3 — Session Day 2 · 1:30pm-1:50pm · Sandbox & Platform Engineering; official schedule)
- 2026 06 30 abhishek bhardwaj from fork to fleet designing an agent sandbox cloud pt2 — From fork() to Fleet: Designing an Agent Sandbox Cloud Pt2; Abhishek Bhardwaj (Day 3 — Session Day 2 · 1:55pm-2:15pm · Sandbox & Platform Engineering; official schedule)
- 2026 06 30 ivan burazin kubernetes is not your sandbox — Kubernetes Is Not Your Sandbox; Ivan Burazin (Day 3 — Session Day 2 · 11:40am-12:00pm · Sandbox & Platform Engineering; official schedule)
- 2026 06 30 kevin orellana 1 000 agent tasks in a sandbox what breaks when llms write and run code — 1,000 Agent Tasks in a Sandbox: What Breaks When LLMs Write and Run Code; Kevin Orellana (Day 3 — Session Day 2 · 2:25pm-2:45pm · Sandbox & Platform Engineering; official schedule)
- 2026 06 30 adam azzam don t build agents build environments — Don’t build agents, build environments; Adam Azzam (Day 3 — Session Day 2 · 10:45am-11:05am · Sandbox & Platform Engineering; official schedule)
- 2026 06 29 matt brockman how i learned to stop worrying and love the sandbox — How I learned to stop worrying and love the sandbox; Matt Brockman (Day 1 — Workshop Day · 11:05am-12:05pm · Workshops Day 1; official schedule)
- 2026 07 01 arun sekhar blast radius zero one command openclaw sandboxes in the cloud — Blast Radius Zero: One‑Command OpenClaw Sandboxes in the Cloud; Arun Sekhar (Day 4 — Session Day 3 · 1:55pm-2:15pm · Track M; official schedule)
- 2026 06 29 derek meegan deploying browser agents at scale — Deploying browser agents at scale; Derek Meegan (Day 2 — Session Day 1 · 1:55pm-2:15pm · Expo Stage 4 SE; official schedule)
- 2026 06 29 ross taylor scaling to long horizons algorithms environments compute — Scaling to Long-Horizons: Algorithms, Environments, Compute; Ross Taylor, Chengxi Taylor (Day 2 — Session Day 1 · 2:25pm-2:45pm · Data Quality; official schedule)
- 2026 07 01 miguel gonz lez fern ndez the art of building verifiers for computer use agents — The Art of Building Verifiers for Computer Use Agents; Miguel González Fernández, Corby Rosset (Day 4 — Session Day 3 · 11:40am-12:00pm · Expo Stage 1 NE; official schedule)
- 2026 07 01 rowan christmas yolo mode safely microvm sandboxes for any agent — YOLO Mode, Safely: microVM Sandboxes for Any Agent; Rowan Christmas (Day 4 — Session Day 3 · 1:30pm-1:50pm · Expo Stage 2 NW; official schedule)
- 2026 06 29 rayan garg rethinking environments for long horizon work — Rethinking Environments for Long Horizon Work; Rayan Garg (Day 2 — Session Day 1 · 11:40am-12:00pm · Data Quality; official schedule)
- 2026 06 29 mahesh sathiamoorthy data and environment curation for post training llms — Data and Environment Curation for Post-training LLMs; Mahesh Sathiamoorthy (Day 2 — Session Day 1 · 3:45pm-4:05pm · Data Quality; official schedule)
- 2026 06 30 tina manghnani from framework to runtime running agents with foundry agent service — From framework to runtime: running agents with Foundry Agent Service; Tina Manghnani, Keiji Kanazawa (Day 3 — Session Day 2 · 10:45am-11:05am · Track M; official schedule)
- 2026 06 30 viren baraiya harnessing agents the durable runtime for dynamic workflows — Harnessing Agents: The Durable Runtime for Dynamic Workflows; Viren Baraiya (Day 3 — Session Day 2 · 11:10am-11:30am · Expo Stage 1 NE; official schedule)
- 2026 07 01 yohei nakajima active graph agent runtime babyagi 4 — Active Graph Agent Runtime (BabyAGI 4); Yohei Nakajima (Day 4 — Session Day 3 · 11:10am-11:30am · Graphs; official schedule)
- 2026 06 29 ang li the autonomous computer full stack infrastructure for computer use agents — The Autonomous Computer: Full-stack Infrastructure for Computer Use Agents; Ang Li (Day 1 — Workshop Day · 4:30pm-5:30pm · Workshops Day 1; official schedule)
- 2026 06 30 philipp schmid why agents should have their own sandbox — Why Agents Should Have Their Own Sandbox; Philipp Schmid (Day 3 — Session Day 2 · 1:30pm-1:50pm · Expo Stage 1; official schedule)
- 2026 06 29 du an lightfoot agents that own their inference building production ai agents on dedicated gpus — Agents That Own Their Inference: Building Production AI Agents on Dedicated GPUs; Du'an Lightfoot (Day 1 — Workshop Day · 9:00am-11:00am · Track 7; related YouTube resource; via youtube wFTVEDYVJT0)
- 2026 06 30 paul klein iv bringing agents onto the world wide web — Bringing agents onto the world wide web; Paul Klein IV (Day 3 — Session Day 2 · 11:40am-12:00pm · Computer Use; official schedule)
Related People
- John Craft
- Abhishek Bhardwaj
- Arun Sekhar
- Tina Manghnani
- Pierluca D'Oro
- Robert Brennan
- Samuel Colvin
- Tushar Jain
- Ivan Burazin
- Kevin Orellana
- Adam Azzam
- Matt Brockman
- Derek Meegan
- Ross Taylor
- Chengxi Taylor
- Miguel González Fernández
- Corby Rosset
- Rowan Christmas
- Rayan Garg
- Mahesh Sathiamoorthy
- Keiji Kanazawa
- Viren Baraiya
- Yohei Nakajima
- Ang Li
Related Companies
- Docker
- Microsoft
- Browserbase
- OpenAI
- Amazon AGI Lab
- Cua
- typedef
- Oxylabs
- Amazon Web Services
- MCP Apps
- Navan
- Uber
- Warp
- Prime Intellect
- Meta
- Yugabyte
- Programma Labs
- OpenHands
Transcript And Resource Support
Transcript-backed resources
- youtube 2IxD9OB3XuQ — Continual Learning for AI Agents: From Failures to Durable Improvements - Soheil Feizi, RELAI
- youtube SKDJo2CopRs — Why Eval++ Is the Next Great Compute Primitive — Sunil Pai & Matt Carey, Cloudflare
- youtube JnubYCYunk8 — Browser Agents Don't Need Better Models. They Need Better Eyes. - Kushan Raj, ARK
- youtube wFTVEDYVJT0 — Building Agents with Amazon Nova Act and MCP - Du'An Lightfoot, Amazon (Full Workshop)
- youtube c 2eEv2ou7Y — Why MCP and ChatGPT Apps Use Double Iframes — Frédéric Barthelet, Alpic
- youtube TNwJ1LMiENk — Stop Making Models Bigger, Make Them Behave — Kobie Crawford, Snorkel
- youtube 4uFVSLgD2Q4 — Agents in Production: How OpenGov Built and Scaled OG Assist - Gabe De Mesa, OpenGov
- youtube YYH0DMQr30A — Task Fidelity Scaling Laws — Kobie Crawdord, Snorkel
- youtube UPwGaM2MKHY — The Log Is The Agent - Ishaan Sehgal, Omnara
- youtube ILdE7FaAjVA — Under 5 minutes to a deployed LLM endpoint — Audry Hsu, RunPod
- youtube _B4Pv9ttFgY — Building Agent Interfaces: Lessons from Chrome DevTools (MCP) for Agents — Michael Hablich, Google
- youtube ghJmWQCIHRM — The agent-ready web: Simplify user actions with WebMCP — Tara Agyemang, Google
- youtube HvZXAOZ3iv8 — What Lies Beneath the API — Benjamin Cowen, Modal
- youtube btxGmN8RvNU — Your Agent's Biggest Lie: "I Searched the Web" — Rafael Levi, Bright Data
- youtube zDGHt0LB dA — GPU Cloud Deployment Without Leaving Your IDE — Audry Hsu, RunPod
- youtube hCMrEfPG2Yg — Beyond Components: Designing Generative UI for MCP Apps — Ruben Casas, Postman
- youtube ObTPqBGsEbA — The Production AI Playbook: Deploying Agents at Enterprise Scale — Sandipan Bhaumik, Databricks
- youtube DqtmZE6Hl0g — The Prompt is the Platform - Dominik Tornow, Resonate HQ