Research readout, updated Oct 15, 2026. I read 65 sources (64 dated) from Oct 2025 to Oct 2026 (podcasts, newsletters and posts by respected product people) and compared what they do with my own agent setup. Every claim links to its source.
Pro PMs have settled on one setup
An agent in a terminal or desktop app, small markdown context files it reads every session, skills for repeat jobs, a scheduled morning briefing, and connectors into Slack, calendar and analytics.
My setup is ahead on the plumbing
Many agents at once with a lead per project, rules enforced by hooks, a weekly pass that turns raw notes into lasting knowledge, phone access, and a fresh reviewer before anything ships.
The gap is judgment
Cagan, Doshi and Anthropic's own staff study warn that AI speeds up output without improving outcomes. The best answer: a critic on your thinking, error reviews of your own AI, and an outcome metric.
Next to try:
Me vs the field
Each row is a practice. The first bar is how many of the respected people I read push it; the second is how far along I am. Click a row for who says so and the sources.
Pros pushing itDoing itPartlyNot yet
PracticePros pushing itMeStatus
The ten practices pros recommend most
Ranked by how many credible people recommend each and how concrete their advice is. Click a practice for its sources.
1
Torres, Gupta (PM OS), Rekhi, JJ Englert, Together AI
Watch out: too much irrelevant context makes it worse (Torres); memory bloats unless pruned (Cat Wu, Peter Yang)
2
Claire Vo, Aman Khan, Colin Matthews, Rohan Varma (OpenAI), Gupta
Watch out: skills "do not eliminate the need for judgment" (Vo)
3
Daniel Blum, Englert, Abhi Muchhal (OpenAI), Torres, Zapier's PM rubric
Watch out: constant system tuning can replace shipping (Wade Foster)