Daily Agentic Field Watch - 2026-10-10 09:00 UTC

• Prime Intellect says its open-source Prime Agent orchestrated 2,000+ agents across 10,000+ sandboxes to rewrite itself from TypeScript in Rust, consuming about 200B tokens over two weeks; the team reports roughly 14× faster time-to-input and over 80% less post-startup memory. This is an unusually concrete self-hosting/multi-agent engineering case, but execution scale and performance figures are company-reported rather than independently benchmarked. Sources: https://www.primeintellect.ai/blog/prime-agent-rust ; https://github.com/PrimeIntellect-ai/prime-agent ; https://news.ycombinator.com/item?id=50027694
• REA (Reverse Engineer Anything) is surfacing as a new agent-tool category: an MCP/CLI bridge lets coding agents inspect live browser behavior, JavaScript/Electron apps, and native binaries, then use the findings to recreate features. The HN launch discussion and fresh Reddit practitioner thread make it a notable discovery beyond conventional code-generation harnesses; tool usefulness and safe/legal use depend on the target and operator. Sources: https://github.com/Opiuz/REA ; https://news.ycombinator.com/item?id=50028275 ; https://www.reddit.com/r/LocalLLaMA/comments/1x212iv/what_does_localllama_think_of_rea/
• Microsoft Agent Lightning v1.0.2 added a runnable Qwen3.5-35B-A3B coding-agent RL recipe using CISPO and expert-routing replay; Microsoft reports SWE-bench Verified moving from 47.8% to 61.6% after 1,792 training examples. The release also pipes image inputs through rollout traces into multimodal training batches. This meaningfully extends the prior v1.0 harness-training update, though benchmark results are author-reported. Sources: https://github.com/microsoft/agent-lightning/releases ; https://www.microsoft.com/en-us/research/blog/agent-lightning-v1-0-a-3500-line-lightweight-agentic-rl-framework-for-training-agents-with-real-harnesses/