AI coding agents

Perplexity has released the sandbox component of its search agent as open source, adding a new entrant to the emerging field of sandbox layers for AI agents. Sandboxes let agents run code and browse safely in isolation, and open-sourcing the tool gives developers a free building block for similar systems. The move is being covered by technology and crypto-adjacent news outlets this week.

Google Chrome team member Paul Irish released a resource called modern-web-guidance, aimed at getting AI coding assistants to stop generating JavaScript and web patterns from around 2019 and instead use current browser primitives and APIs. The project, hosted on GitHub, provides guidance that developers can feed to their agents so generated code reflects modern web capabilities rather than outdated defaults.

A new open-source tool called Klad has been released, presenting itself as a paint application that can be driven by coding agents such as AI assistants. The app is available as source-available code on GitHub, letting developers inspect and adapt it. Early discussion online is focused on how AI agents could automate creative and design workflows, not just software tasks.

A widely shared commentary from developer Theo (t3.gg) argues that AI coding agents have historically performed poorly at building and handling user interfaces, while suggesting that something has changed as of today to address that weakness. The claim is drawing large attention from software developers, with many debating whether agents are finally becoming capable at front-end and UI work.

Developer Julius Brussee released 'Caveman', a Go-based skill and proxy for coding agents that rewrites the model's output in stripped-down, simplified language to cut token usage by roughly 65%. The tool's promotional pitch leans on the joke 'why use many token when few token do trick'. It is circulating as a quirky but practical way to reduce costs when running AI coding assistants.

Cloudflare has published an open-source coding-agent skill for carrying out multi-phase security audits. The tool produces independently verified, machine-readable findings, letting AI coding agents run structured security reviews of codebases. Its release is drawing attention from developers interested in integrating automated security auditing into AI-assisted development workflows.

Y Combinator president Garry Tan says he spends as much as $200,000 a year on AI tokens, reflecting how heavily he relies on AI agents in his daily work. The figure highlights how expensive intensive use of AI coding and automation tools can be, even for a single executive, and has drawn discussion about the cost and productivity trade-offs of running many agents in parallel.

A detailed write-up describes how a real company runs 17 Claude Code bots in production for roughly $100 per month, breaking down actual costs and operational experience. The account of how cheaply AI coding agents can be deployed at scale is drawing attention among developers weighing the economics of AI-assisted engineering.