HumanExperiences

AHE: A Self-Improving AI Co-Worker Platform AHE is an orchestration platform for long-running autonomous work. Give it a goal it plans, acts on a real Mac, reviews its own work, and gets better every round. Governance, not just execution. AHE has a four-level hierarchy: Strategy → Unit of Work → Job → Task. Each level reasons independently, reading only compressed summaries from the level below. That's what makes week-long autonomy cheap and coherent. Multiple agents work as one co-worker, sharing what they learn. Self-improvement is the core. After every round, an evaluator scores each Job against real evidence not summaries, but the actual output, the actual metrics, opened on the real screen. When evidence is missing, it dispatches a real Job to go get it: read the metrics, check the file, whether the work lives online or on the local machine. Those scores feed the next round's planning: keep what worked, fix what didn't. Judgement accumulates. The agents builds a model of what works, not just a log of what happened. Any Mac, any app. AHE operates macOS directly with mouse, keyboard control and screenshots. Figma, Canva, the browser, anything on screen. One Mac runs one Workflow. And it can be Scaled by adding Macs. Safety is structural. Hard walls block system settings, purchases, account changes. CAPTCHAs are reported, never solved. Credentials stay scoped to their site. Built for the Meta Hackathon. Meta Muse powers every layer, planning, judgement, and execution with one API key.