LLenny's Podcast
← All frameworks
ProductivityThe future of software development with OpenAI’s Sherwin Wu

Hand the Toil to the Model First

Delegate the boring, repetitive parts of engineering to agents before anything else, and collapse the path to production.

Difficulty
Easy
Time to result
~weeks to results
Steps
3
Confidence
85%

The highest-return first move with coding agents is to hand them the parts of the job that annoy people most — code review, lint fixes, CI, and the deploy path — rather than the creative work. This both frees engineers for the fun, high-judgment work and dramatically shrinks the friction between writing code and shipping it, which is why AI-heavy engineers open far more PRs.

Origin

Sherwin Wu, describing how OpenAI internally automated code review and the CI/deploy pipeline with Codex.

Core principles

  • 01The tasks worth handing to models first are the ones that annoy you or are most boring
  • 02A smart second pair of eyes (the agent) can replace much of the manual review load
  • 03The bottleneck after writing code is getting it into production through tests, lint, and review
  • 04Automating toil compounds into far more shipped PRs

How to run it

  1. 1

    Route all PRs through agent review

    Have the agent review 100% of PRs so every merge gets automated suggestions before a human looks.

    Pro tip Steer the reviewer in the right direction; for small PRs the author reviewing the agent's output can replace a second human reviewer.

    Watch out Stay thoughtful about which PRs go fully agent-reviewed — don't let attention drop to zero on everything ('don't let the brooms go crazy').

  2. 2

    Automate the path to production

    Build tooling so that lint errors, tests, and the post-push deployment steps are handled by the agent — e.g. auto-patch a lint error and restart CI.

    Pro tip Collapse these steps into as little engineer work as possible so shipping is near-frictionless.

  3. 3

    Reclaim the freed time for high-judgment work

    Redirect the engineer's attention from toil to the creative, exploratory parts of the job the model isn't handling.

In the wild

From 20-minute to 3-minute reviews

Wu, who hated code review since owning the constantly-touched Quora newsfeed code and facing 20-50 reviews each morning, now has Codex review all PRs with suggestions pre-baked. GPT-5.2 in particular became strongly adept at reviewing code when steered.

Code reviews dropped from a 10-15 minute task to sometimes 2-3 minutes, and engineers using Codex open 70% more PRs — a gap that keeps widening.

Common mistakes

Automating creative work first instead of toil

Handing the model the open-ended, high-judgment work first squanders its best early ROI; the annoying, repeatable parts (review, lint, CI) are where automation pays off fastest and frees humans for the fun work.

Dropping human attention to zero everywhere

Trusting the agent to review and merge with no human oversight on any PR risks letting errors compound — the 'brooms go crazy' failure — so calibrate which PRs are safe for full automation.

Is it for you?

Best for

Engineering teams adopting coding agents who want an immediate productivity win

Not ideal for

High-stakes changes where every PR genuinely needs careful human review before merge

From the transcript

the things that tend to we hand we tend to hand to the models immediately tend to be the things that annoy us or like…

15:30

Codex reviews 100% of all of our PRs at this point

15:00

it makes you know code reviews go from a you know I don't know 10 15 minute task to sometimes even just like a two…

16:30

they're actually opening 70% more PRs

04:30

From the episode

“Engineers are becoming sorcerers”

The future of software development with OpenAI’s Sherwin Wu