PersonalAgentBench: Dots first, Hark second, Grok Bot tied fourth
micro1's trust-focused leaderboard ranks six personal agents; Hark trails only OpenAI's Dots, and Brett Adcock says he wants first place.
What you can do now: if you are choosing a personal agent, micro1's PersonalAgentBench now ranks six of them on the same 20 questions. Hark is the runner-up, scoring 75% on getting tasks done and 70% on doing them without breaking your rules. Brett Adcock quoted the result: "How embarrassing. Will work to crack #1."
What changed#
On Oct 12, micro1 added Hark and OpenAI's Dots to PersonalAgentBench and doubled its task set. At launch on Oct 6 it tested four agents: Instinct, Muse, Grok Bot and Gemini Spark. The full provisional leaderboard, as published on micro1's page (checked Oct 13, 6:20 AM ET):
| Rank | Assistant | Task completion | Trusted completion |
|---|---|---|---|
| 1 | Dots | 80% | 80% |
| 2 | Hark | 75% | 70% |
| 3 | Gemini Spark | 70% | 65% |
| T-4 | Instinct | 65% | 65% |
| T-4 | Grok Bot | 65% | 65% |
| 6 | Muse | 55% | 55% |
Across the six, mean Task Completion was 68.3% and mean Trusted Completion 66.7%. Dots, the leader, are OpenAI's always-on ChatGPT agents announced Sep 29, each with its own cloud computer and reachable from Slack and Teams (our explainer). micro1 says Grok Bot did better on straightforward multimodal and communication tasks than on prioritization, grounding and account-state work.
What it measures#
Each run gets two pass/fail verdicts. Task Completion means "every required outcome was verified"; a run without enough evidence fails. "Trusted completion" fails a run for a rule violation, an unintended account change, or a false claim of success. micro1 says Hark "was strongest on stateful action and authority," and weaker on "evidence completeness and some personal-context judgments." In plain terms, Hark acts carefully on your accounts, but check that it actually finished and that it understood what matters to you.
How to use it#
Users spent the day building Home panels. Adcock shared one pulling Whoop data and said any data in your account can feed a custom panel. One user's tip: give Hark an image of the widget you want and describe it.
- Open Hark and ask for a panel.
- Attach a screenshot of a layout you like.
- Ask it to show its sources so you can verify the numbers.
Make a panel for my Home screen that shows last night's sleep score from WHOOP and my next three calendar events. Use the attached image as the layout.
Good to know#
- micro1 calls this a "small, Google Workspace-heavy sample," and prompts and accounts were not identical across agents.
- One day-one reviewer listed missing pieces: no macOS app, no two-way voice, no way to take over the browser, no list of plugins. Adcock replied, "Working on the feedback improvements."
- A user said captchas stop Hark from creating accounts; Adcock replied, "Working on it."
- iOS version 2.2.1 shipped Oct 12 with only "Refinements to Hark" as notes.
What to watch#
micro1 says new releases will add tasks and "controlled repeat testing in the coming days." A rerun shows whether Hark closes the 5-point gap to Dots.
