We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Big Ideas
The agent moat is moving from initial build to accountable operation. Businesses want ordinary mess removed—email triage, call memory, CRM hygiene, invoices, search, lead routing, and reporting—but value appears when an agent holds context across systems and carries work forward. The PM problem is therefore operating design: define what the agent can see and use, when it may act, how output is checked, and what happens under uncertainty.
Platforms are already absorbing bespoke agent work into native surfaces, so horizontal “we make AI agents” offers are exposed. More durable bets sit in vertical workflows, cross-system integration, private/local deployments, evaluation-heavy systems, and high-cost-of-error operations; maintenance is part of the product because an agent can stay online while silently degrading. Roadmaps should budget for evaluation, monitoring, correction, permissions review, and redesign—not just launch.
Tactical Playbook
Make the prototype the shared decision object. One PM reports that rough, unbranded wireframes improved collaboration with designers and engineers, sped alignment, and enabled collective ideation; the tradeoff was that written requirements became more painful and agreements moved into concept-building. Use low-fidelity prototypes early, label assumptions and roughness explicitly, then capture final decisions once the concept stabilizes so speed does not erase traceability.
Case Studies & Lessons
A freemium test rejected forced conversion. A student productivity app split new users among the existing free plan, a 14-day Premium trial followed by read-only access, and an 80-hour usage paywall. The normal/free experience performed best, so the founder kept the core app free and shifted to contextual Premium prompts after repeated use of a relevant feature, alongside multiple price points and regional pricing.
The results are directional, not proof of causal lift: registered users rose from 3,080 to 5,561, monthly active users reached about 1,500, trials rose from 2 to 59, five converted—roughly 8.5% of a small sample—and revenue reached about €185. Most growth came organically through Google, while infrastructure costs came under control. The PM lesson is to test whether the problem is gating, value, or positioning before degrading the free product. A useful diagnostic is a non-leading question such as “What’s the biggest value you get?” Answers about free features suggest a premium-value problem; answers about premium features suggest a positioning problem.
Career Corner
High-talent hiring is a targeting problem, not a volume problem. Cursor’s head of talent calls the conventional 100-outreach/20-replies funnel “remainder” hiring. His alternative: define “great” by stack-ranking skills and experiences, explain the role’s impact and success criteria, map a finite target list—50 is his example—and pursue it. Use referral questions tied to a specific trait, such as collaboration with designers, rather than “who’s the best?”
For PM candidates, build a work sample that demonstrates judgment and execution, not just polished artifacts. The interview cites work samples as the strongest predictor of success, and Cursor uses project-based, side-by-side on-sites because removing work trials weakened its signal.
Tools & Resources
Simplify AI harnesses as models improve. Aakash Gupta’s Claude Code note argues for “outcomes > steps”: specify the desired outcome, format, quality bar, examples, and guardrails rather than a large procedural prompt; move context into skills and libraries, and maintain the harness because 1–2% gains on each task compound. The note also links a free harness-upgrade skill. For PM workflows, keep the brief outcome-led, put reusable context in a maintained skill, and evaluate outputs against a small explicit quality bar.