21. The honest comparison
📣 Real-user report — shared publicly by a Muse user, not independently verified.
The job: One user ran Muse side-by-side against another agent and reported a higher success rate — praising the working-state indicators, approvals UI, and memory. His caveat: CAPTCHAs and logins can still need manual takeover.
I want to stress-test you on a real task: [organize last week's inbox into action items]. Narrate what you're doing as you go, ask approval before anything irreversible, and at the end give me an honest report: what worked, what you couldn't do, and where you needed me. No sugar-coating.
🔌 Needs: Gmail — connect in the Muse app first (setup guide).
What Muse will do:
- Take on one real task, like organizing last week's inbox
- Narrate what it is doing as it goes
- Ask approval before anything irreversible
- Report honestly what worked and what it couldn't do
- Flag every point where it needed your help
Why it works: turn evaluation into a delegation — the "no sugar-coating" report tells you exactly where the agent's limits are before you trust it with bigger jobs.
Common failures:
- Vague tasks produce vague reports — pick one concrete, real task to test
- Skipping narration hides the reasoning — ask it to narrate as it goes
- No approval rule risks irreversible actions — require approval first
Prompt variations:
- Stress-test with calendar triage instead of inbox cleanup
- Run the same task on two agents and compare the honest reports
- Ask for a difficulty score per subtask to spot automation candidates
Reported by @sidwyn on Threads · Sep 2026
Loading comments…