Ideas Radar: March 27, 2026

A real browser testing framework purpose-built for AI agents. Current agents write code but have no way to verify it works visually. A tool that lets agents run QA in a real browser, record video of bugs found, and iterate until passing would eliminate the biggest gap in the vibe coding workflow.

An iMessage/SMS bridge for AI coding agents. The ability to text your agent from your phone and get results back changes the entire interaction model. Currently only available as a Claude Code plugin — the first open-source version that works with any agent framework would be immediately valuable.

Enterprise-grade permission classifiers for autonomous agents. Auto mode showed that binary permissions (approve all vs manual) can be replaced by ML classifiers that evaluate each action. Building this as a standalone service that any agent framework can plug into would solve a universal problem.

A CVE tracker specifically for AI-generated code vulnerabilities. With Georgia Tech scanning 50K+ advisories and finding dozens of confirmed AI-introduced bugs, a real-time dashboard that tracks which AI tools produce which vulnerability classes would help teams make informed tooling decisions.

A prediction market automation framework. One trader reportedly turned $200 into $3.7M using OpenClaw for probability harvesting on Polymarket. The demand for accessible, no-code market-making bots that handle probability math automatically is clearly massive.

AWS architecture skills as a standalone product. Agent Plugins for AWS just launched inside Claude Code, but most teams use multiple AI tools. A framework-agnostic AWS architecture agent that works with any LLM would capture the much larger market.

An open-source alternative to RunClaw that adds the missing pieces to OpenClaw without the vendor lock-in. RunClaw charges $1/month to wrap OpenClaw. The gap they identified (setup complexity, missing website/video/slides capabilities) is real — solving it in open source would be more defensible.

A research agent that can replicate experiments and simulate peer review. Feynman showed this is possible — give it a question, get a cited meta-analysis back. The academic research vertical for agentic AI is wide open.

An AI agent monitoring dashboard with pixel-art visualization. Someone built one for OpenClaw where agents walk around a virtual office based on their activity. Whimsical but genuinely useful for understanding multi-agent systems at a glance.

A semantic search layer for personal notes. 2,649 notes in Apple Notes with zero way to search by meaning. Google Photos proved people love semantic search for images — the same thing for all personal content is a massive opportunity.