Addy Osmani argues that if the AI layer becomes good at anything, it will excel where there is an answer key. Traditional schooling anchored success around getting the right answer; in the age of AI, durable human value is less about solving problems and more about choosing which problems to solve and judging the quality of solutions. Increasingly, agents will do the rest faster and better.
Author background
Addy Osmani began engineering at 16, building a browser in rural Ireland. He worked at Google for over 14 years, leading engineering teams on Chrome, Gemini, and Cloud AI, and has written several O’Reilly books. He says he has declined offers from frontier labs and FAANG companies when the fit was not right.
Optimize for scarce resources
Most career advice still holds: get on the rocket ship and avoid over-optimizing your seat. What has changed is which resources are scarce. Capital and time feel abundant; reputation, real relationships, and a public track record of good work remain rare. Osmani’s plan: do good work and make sure people who value good work see it. In a world where quick “vibe coding” can make money easily, those quick wins have little value compared to choosing meaningful things to ship.
Learn to find problems, not just solve them
Where LeetCode once measured skill, agents now make many such problems cheap to solve. Selection—finding the important problems—has become the scarce skill. Osmani recounts that he often noticed a performance pain point (dial‑up slowness), invented chunked multi‑connection fetching, then moved on when he realized he wouldn’t spend his life on that problem. Finding problems preceded solving them.
He’s seen top students fail when an agent breezes through their problem set. The difference is judgement and intuition: the strong bring those; the rest bring a prompt. Taste is pattern matching, but that pattern matching must be earned by doing the work. The real risk is not agents writing bad code (we’ve seen that before); it’s losing the ability to tell when output is correct. Judgment will atrophy if practitioners rely on agents for every repetitive rep.
Recommendations to preserve and build judgment:
- Pick a few important problems and do them the hard way, without agents, to build deep mental models of systems and languages.
- Read far more code than you write.
- Treat every agent diff like a human review that you must carefully justify.
- Keep a private log of agent suggestions that look wrong and flag them confidently — that is where taste accumulates.
Shift from doing to directing
You must learn to delegate to agents the way you would to a person: scope the task, define done, calibrate trust, and verify the result. Autonomy is a per‑task setting, not a rank. Turn it up on small, reversible, cheap‑to‑check tasks; turn it down where mistakes are hard to undo.
Specification and verification are distinct but complementary skills. An agent is not as good as the intent you hand it; the best engineers write precise specs. Clear thinking made legible matters. Verification should not be an agent grading its own homework; delegation at scale without verification is demoralizing.
Own what you ship
If an agent wrote the change and it breaks in production, “the AI did it” is not a defense. Your name is on the change: adopt an accountable posture, understand what went out, and know how to fix it.
Solve the most ambitious version of the problem
Citing Rich Sutton’s “bitter lesson,” Osmani notes that general methods that scale with compute often outperform hand‑tuned equivalents. As a career lesson, solving the easy version of a problem yields little value; the value concentrates in the hard version.
Sprint the last mile
No turnkey agent yet writes whole systems end‑to‑end. Typically an agent will produce about 70% of a feature quickly; the last 30% — debugging edge cases, choosing architecture, cultivating taste — is where the real value lies. When first drafts are effectively free, finish is the product.
Osmani’s tactic: every few months he rebuilds projects from scratch using the latest cutting‑edge model. He finds this less exhausting than nursing half‑hearted old code back to health. His job as an engineer, he says, has been to finish strong; the difference between finishing strong and finishing okay is polish — often an extra hour that shows to everyone who matters.
xG, reputation, and finishing
Using the xG (expected goals) analogy: xG measures how many chances your play should produce; finishing measures conversion. Over a career you can influence both how many chances you get and how well you finish them. For Osmani, large opportunities came from public work, not from jobs he applied to. You cannot script which chances arrive, only whether you stand where they land. Bet on the team and the market opportunity, not merely the product demo; the demo is a snapshot, the team is the trajectory.
On superintelligence and the near future
Osmani believes future models might replace much of knowledge work, though not overnight and not everything. New jobs will appear, and verification will remain a bottleneck. Someone must decide which problems are worth solving and allocate judgement accordingly — that can be you. The gate to AI research is smaller than it seems: you don’t need a lab to build intuition — use models heavily and turn observations into evaluations. Evals and benchmarks are where understanding lives.
Conclusion
The world is not short on opportunity; it is short on people who can find the right problems, judge whether the machine solved them, and finish past where the machine stopped. In agentic systems the last few feet can be effectively infinite because agents scale output; human attention does not. Protect your attention: anything gradable by someone else will be automated. The career is the ungradable part — choosing what matters, honestly judging when you’ve got it, and answering for it. Do that in public, near the hard problems; the rest tends to follow.
This piece grew out of Phil Chen’s original writing, which Osmani recommends reading. He also notes AI Codecon: Building with Open Source AI, a free half‑day virtual conference on August 31, featuring developers and experts working with open‑weight models, self‑hosted infrastructure, and real‑world AI workflows.
