Hark released Handoff², the second version of its computer-use agent embedded in Hark Pro, reporting an average 2x end-to-end speedup over the original Handoff across web tasks such as grocery ordering, flight booking, and warranty forms. Gains stem from a co-evolved agent harness and reinforcement learning (RL) recipe that penalized wasted turns, unnecessary thinking tokens, and latency. The new harness permits the model to emit more actions per turn; RL training induced three behaviors: batching actions like multi-field form fills into single turns, grouping repeated patterns into reusable composite actions, and reusing those composites throughout a session. Each request runs on a dedicated virtual computer with its own browser, filesystem, and terminal, with the model outputting cursor and keyboard actions targeted at screen coordinates. On Hark's internal 478-task browsing benchmark spanning 238 websites, Handoff² achieves 82% completion versus 73% for v1, with 46% less time per task, though it trails a Claude Opus 5.5 (max) baseline by up to 5 points in the same harness. Alignment scores rose to 93.8% from 90.3%. Speedups scale with task length: flight booking up to 3.8x faster and long research over 5x faster, reducing a 45-minute task to roughly 10 minutes. All figures are self-reported; no third-party or Online-Mind2Web results for Handoff² have been published. Handoff² is accessible via web, iOS, and Android under Hark Pro tiers: free, $20/month for 2x usage, and $100/month for 10x usage. No public API or model weights have been announced.
Read original
marktechpost