I benchmarked Hark, Instinct and Muse on 6 tasks — only 2 points separated them
AgentWorkinc · reddit · 2026-10-16
The author ran identical 6 tasks — logic puzzle, sourced research, Python function with tests, sub-$20 hardware project, one-page AI workflow guide, live browser task — through Hark, Instinct and Muse with a 10-point rubric each.
Scores: Hark 57/60, Instinct 55/60, Muse 55/60. Key findings:
- Muse was the best builder: 10/10 on coding and engineering; its hardware concept "Leash" was a physical approve/deny terminal for sensitive agent actions. Weakest at research (7/10), mostly tool selection.
- Hark was the most consistent: perfect on logic, coding and the live browser task; proposed "AgentMeter", a gadget measuring agent hardware power draw under $20. Nothing below 9/10.
- Instinct aced logic and the guide (covering delegation, cost-saving, multi-agent workflows, verification); weakest round was engineering (8/10).
Caveats: one run per task, moderate prompts, human grading with GPT assistance — treat numbers as rough ordering. Takeaway: on moderate tasks, the hype gap between agents doesn't show up in performance. Round 2 will add timed speed and failure-recovery tests.
More from Apps
- ChatGPT UI screenshot reveals Wallet coming soon plus Email, Calendar, Health and Finances integrations — jxnlco · 2026-10-16
- Law student builds all-in-one AI study platform LexLab Studio with ChatGPT and Muse — Acrobatic_Rhubarb598 · 2026-10-16
- Only 4.5% of US consumers pay for AI, but the top 1% of spenders drive 19.5% of revenue — dumpshoot · 2026-10-16
- Australian e-commerce brand AI bots look identical, likely from one vendor — gaganghotra_ · 2026-10-16
- Small business owner gives his AI assistant Dot its own phone number to handle clients — Early-Protection2386 · 2026-10-16
- Obsidian plugin Qiaomu AI RSS adds text-to-speech with karaoke highlighting — vista8 · 2026-10-16