LLenny's Podcast
← All frameworks
ProductivityMike Krieger (co-founder of Instagram)

Hunting New Bottlenecks When AI Writes the Code

When AI removes the coding bottleneck, constraints shift up and downstream — go find them.

Difficulty
Advanced
Time to result
~months to results
Steps
5
Confidence
90%

Once AI writes the bulk of your code, the historical bottleneck (engineering build time) dissolves and new constraints appear elsewhere in the system. Applying Theory-of-Constraints thinking, Krieger maps them into three zones: an upstream decision/alignment bottleneck, a midstream coordination bottleneck, and a downstream merge/review/launch bottleneck. The job becomes continuously finding and re-architecting whatever the new binding constraint is.

Origin

Mike Krieger explicitly invokes 'The Goal' and its critical-path / Theory of Constraints thinking (Eliyahu Goldratt) to describe how bottlenecks migrated at Anthropic once ~90% of code became AI-written.

Core principles

  • 01Removing one bottleneck simply relocates the constraint elsewhere
  • 02Upstream: deciding what to build and aligning people becomes the new limiter
  • 03Midstream: preventing collisions and pre-handling edge cases so building isn't blocked
  • 04Downstream: merge queues, review, and launch coordination get overwhelmed by volume
  • 05Existing org assumptions and roles lag the new reality and must be revisited

How to run it

  1. 1

    Accept the old bottleneck is gone

    Recognise that engineering build time is no longer the constraint when most code is AI-generated; volume of code and pull requests explodes.

    Watch out Holding onto old process assumptions ('more headcount') often masks the real new constraint.

  2. 2

    Fix the upstream decision bottleneck

    Provide a 'minimum viable strategy' — just enough direction to empower people to prototype and build at the edge of model capability without waiting for alignment.

    Pro tip Krieger admits he hasn't fully solved this yet; treat it as an ongoing tuning problem.

    Watch out Too little strategy leaves teams unaligned; too much strategy chokes their autonomy.

  3. 3

    Handle the midstream coordination bottleneck

    Put in place ways to avoid stepping on each other's toes and to think through edge cases ahead of time so engineering isn't blocked mid-flight.

  4. 4

    Re-architect the downstream merge, review, and launch bottleneck

    Rebuild the merge queue for far higher PR volume, and rethink review — e.g. use a second model to review and have humans do acceptance testing rather than line-by-line review.

    Pro tip The Claude Code team stopped line-by-line reviewing AI PRs and moved to model-review plus human acceptance testing.

    Watch out Unchecked, this risks an unmaintainable codebase, so keep a human acceptance gate.

  5. 5

    Repeat as constraints keep moving

    Treat bottleneck-hunting as continuous; expect the way you conceive of building and shipping to change substantially year over year.

In the wild

Anthropic's merge queue blowout

With AI generating over half (and rising toward 70%+) of pull requests, so many more PRs were submitted that the merge queue 'completely blew out the expectations' and had to be fully rearchitected.

Anthropic re-engineered the merge queue and shifted review processes, exposing decision-making, coordination, and launch as the next bottlenecks.

Common mistakes

Reflexively adding headcount

Krieger admits assuming 'we need more headcount' when the real need was a few founder-type engineers and re-architected process — throwing people at a relocated bottleneck doesn't clear it.

Keeping line-by-line review on AI-generated PRs

AI produces PRs larger than humans can meaningfully review line-by-line, so clinging to the old review ritual becomes its own bottleneck; shift to model-review plus acceptance testing.

Is it for you?

Best for

Engineering and product leaders whose teams are rapidly adopting AI codegen and are hitting unexpected process bottlenecks.

Not ideal for

Small teams or early projects where AI is not yet writing a large share of code and the classic build bottleneck still dominates.

From the transcript

I don't know if you've ever read is it the goal the classic like process optimization book and you realize there's like this like critical…

14:00

You know there's an upstream bottleneck which is decision-m and alignment

14:30

we had to completely rearchitect it because so much more code was being written and so many more pull requests were being submitted that it…

14:00

can you use a different cloud to review it and then do the human almost like acceptance testing more than trying to like review line…

16:00

From the episode

Anthropic’s CPO on what comes next

Mike Krieger (co-founder of Instagram)