The Radar

vLLM's Micro-Agent approach uses multiple collaborating agents within a single model API call to outperform larger frontier models. The idea is coordinated inference rather than brute-force scale. It is an open-source release from the vLLM team and already generating solid discussion.

DoorDash open-sourced a terminal UI built internally for managing coding agents that run long tasks. It addresses a practical gap in the current agent tooling ecosystem, most tools assume short bursts, not extended autonomous runs. Clean open-source release from a real engineering team.

VibeRaven adds workflow structure around AI coding agents to make them production-ready. It targets the gap between a demo agent and one you can actually rely on in a real codebase. Early project but pointing at a genuine pain point.

Exfault automates mobile app security testing using an agentic approach. It is aimed at developers and security teams who need to run penetration tests without deep specialist knowledge at every step. Niche but well-timed given how little AI tooling exists in mobile security.

Reference MCP is a Model Context Protocol server that indexes past agent sessions and makes them searchable by other agents. It is a lightweight approach to giving multi-agent systems shared memory without a heavy infrastructure setup. Useful for anyone building persistent agent pipelines.