Blog
Agent safety, isolation, and building with Islo.
· Assaf Ben Josef
We automated implementation first and it backfired. Starting with review and verification instead, and wiring it all together with webhooks and label-driven triggers, gave us an autonomous engineering loop where a Linear label goes in and a merged PR comes out.
· Rotem Tamir
Introducing RewardHackBench, an open-source benchmark for measuring whether sandbox policies prevent reward hacking in agent evaluations.
· Adam Gold
A technical look at what Agent Sandbox solves inside Kubernetes, what it leaves to the platform team, and how to decide if it's the right primitive for your setup.
· Adam Gold
Fast boot times are useful for request-shaped compute. Coding agents need something closer to a secure, persistent, workstation-shaped runtime.
· Adam Gold
A framework for evaluating what you need to run AI coding agents at scale and where most teams get stuck.
· Adam Gold
Claude Code for Web is great for quick coding help in the browser. Islo lets you run Claude Code alongside Cursor, orchestrate multiple agents, and connect to real services - all unattended.
· Adam Gold
Cursor Cloud Agents is great for quick edits in the browser. Islo lets you run Cursor alongside Claude Code, orchestrate multiple agents, and connect to real services - all unattended.
· Adam Gold
Stateful fake servers for GitHub, Slack, and Stripe that work with official SDKs - built for AI agent development.
· Yossi Eliaz
Monty sandboxes runtime-generated Python. Islo gives coding agents a zero-setup environment. Here's what Monty blocks, what can still crash, and how the layers fit.
· Adam Gold
Here are five sandbox options that work with Claude Code (or with AI coding agents in general), and how they compare.
· Adam Gold
Why sandboxing matters for AI agents - isolation, boundaries, and blast radius so you can run agents unattended without risking credentials or production.
· Adam Gold
How I sent my own database credentials to a server I control using Clawdbot - and why prompt injection and full machine access make that possible in about 20 minutes.
