@nate_512WHAT?! GOOGLE JUST OPEN-SOURCED SELF-IMPROVING AI AGENTS 🤯 It's called Harness², and it lets agents rewrite their own harnesses to get better at new tasks. 100% open source. Here's how it works: 1. Propose changes to prompts, tools, memory, skills, and other components 2. Probe each change by running the actual task 3. Compose the best changes into an improved harness No retraining. No ground-truth answers. Model weights stay frozen. The clever part? Each task leaves behind improvements the agent can reuse. By task 50, it's building on lessons from the previous 49. Google reports gains of up to 11.8 percentage points after just one refinement step on a legal benchmark. It supports parallel and sequential optimization, with a local demo requiring no API credentials. FREE and open-source. GitHub repo in 🧵↓

@nate_512WHAT?! GOOGLE JUST OPEN-SOURCED SELF-IMPROVING AI AGENTS 🤯 It's called Harness², and it lets agents rewrite their own harnesses to get better at new tasks. 100% open source. Here's how it works: 1. Propose changes to prompts, tools, memory, skills, and other components 2. Probe each change by running the actual task 3. Compose the best changes into an improved harness No retraining. No ground-truth answers. Model weights stay frozen. The clever part? Each task leaves behind improvements the agent can reuse. By task 50, it's building on lessons from the previous 49. Google reports gains of up to 11.8 percentage points after just one refinement step on a legal benchmark. It supports parallel and sequential optimization, with a local demo requiring no API credentials. FREE and open-source. GitHub repo in 🧵↓