The Rise of
the Product Engineer
When code costs nothing, what's left of the job?
"When it comes to shipping AI products, you want engineers, not researchers."
- swyx, "The Rise of the AI Engineer", June 2023
The bottleneck moved,
and it's going to keep moving
We're going to go fast
- Every number in the next few minutes has its source on the slide
Coding 8x faster to ship 1.3x faster
- NBER, 100,000+ GitHub developers: agents mean +741% code, +30% releases
- LinearB, 8M pull requests: agent PRs wait 4.6x longer for a reviewer
- Cisco / SmartBear: review accuracy collapses past ~450 lines an hour
- Faros, 22,000 developers: time in review +440%, incidents per PR +240%, merged with no review +31%
- Anthropic: 80%+ of merged production code is written by Claude
The cost of producing
plausible code has collapsed.
The cost of knowing
whether to trust it has not.
The machines are taking over review
- GitHub Copilot: more than 1 in 5 code reviews on GitHub. Since 1 Sept, its review can count as an approval.
- "Humans may review pull requests, but aren't required to." - OpenAI
- "The End of Code Review" - Martin Monperrus, KTH, June 2026
- 4 AI reviewers, same 146 PRs: 617 lines flagged, 0 flagged by all four
- LLM reviewers hit the same cliff as human ones, on smaller diffs
Taking the human out entirely backfires
- "I was wrong. Please, please read the code." - Dex Horthy, HumanLayer
- Amazon, March 2026: AI-assisted changes from junior and mid-level engineers now need senior sign-off
- Anthropic postmortem: bugs got past "multiple human and automated code reviews, as well as unit tests, end-to-end tests, automated verification, and dogfooding"
Agents can pass 80% of a coding benchmark's tests
but only 54% if those same PRs are judged by a maintainer
Fix one bottleneck and you get another one
- Eliyahu Goldratt, "The Goal", 1984
Two big assumptions
- 1. Agents are going to eat the entire software development lifecycle
- 2. There is no upper bound to how much software we need
What agents eat, and in what order
- Collapsed: writing code
- Going soon: reviewing and maintaining code
- Next on the chopping block: shipping to production, scaling
- Possibly safe: deciding what to build, defining "good", making it delightful
Only the baker knows what the bakery needs
There's no economy of scale in product decisions
Brooks said the hard part
was never the code
- Fred Brooks, "No Silver Bullet", 1986: essential vs. accidental complexity
"Whoever writes the spec, be it a PM, a lawmaker, an engineer, a marketer, is now the programmer."
- Sean Grove, OpenAI, AI Engineer World's Fair 2025
Whoever defines "good" trains next year's models
- Cognition defined "mergeable", then shipped SWE-2, a model trained to score on it
- Cursor's Bugbot turns your team's thumbs-down into rules: 44,000 learned rules across 110,000 repos
"For the first time in my life, managers are proposing having twice as many PMs as engineers."
- Andrew Ng, July 2025
We've had this job before,
under two names
- 1960s: the systems analyst
- 1980s onward: product manager and engineer, two separate careers
- Now: collapsing back into one job
The product engineer
Finds out what people need,
defines "good", and ships it
- "90% of my skills just went to $0. The other 10% went up 1,000x." - Kent Beck
The market is already paying for it
- Forward deployed engineer postings: +800% in 9 months of 2025
- ~1,000 live postings across 462 companies: OpenAI, Anthropic, Databricks, Stripe, Google Cloud...
- Average total comp ~$240,000. Senior roles clear $600,000.
- Also posted as: solutions engineer, deployment engineer, applied AI engineer, implementation engineer...
"But that's not programming!"
All juniors did was write
the code you told them to
Code review was the apprenticeship,
and we just automated it
Nobody trains
product engineers on purpose
- Google APM program: ~50 people a year, from ~12,000 applicants
So grow them on purpose
- Put juniors in front of customers on day one, not in the ticket queue
- Have them write the spec and the eval; let the agent write the code
- Review their spec and their transcript, not just their diff
- A $240,000 forward deployed engineer is cheaper to grow than to buy
The craft as paid work is mostly over,
and that is a real loss
Ten ugly years.
Then a bigger profession
What to do on Monday
- Read: your blast-radius list. Public APIs, auth, schemas, anything a human has to sign their name to.
- Delegate: anything cheap to verify. Run two reviewers and tell them to assume the code is wrong.
- Define: specs, rubrics, and evals. That's where your best people's judgment goes now.
Measure what ships, not what gets written
- Not: lines of code, pull requests opened
- Instead: what ships, what reviewers accept, the gap between "tests pass" and "would merge"
- The last reviewer is production
We are all product engineers now
Thank you
BlueSky 🦋 @seldo.com
LinkedIn: Laurie Voss