What's next?
AI has moved the bottleneck
March 12, 2026
let's rewind
Claude Code: research preview
the GA to take place a few months after
OpenClawdMoldBot
wait, what?
Plan mode doesn't exist
Cursor to ship it quarters later
GPT 4o, Sonnet 3.5
are the best in class models
February 2025
was 4 years ago
. . .
I'm not getting paid by Pepsi. I prefer Coca-Cola.
What has it
changed for us?
and no, they're not getting smaller
- continous PRs/eng records
- avg PR size has increased (!)
- we are still learning how to use AI
- safe to assume: new releases
the change in productivity and throughput is real
Code shipped
≠
value created
...but it does help creating value,
and we'll get back to it
Problem definition
↓
Solution design
↓
Code creation
↓
Review
↓
Deploy
↓
Observability (Product & Tech)
↓
Learn & iterate
Systems like these don’t like changes.
When one stage accelerates, the constraint moves.
Problem definition
↓
Solution design
↓
Code creation
↓
Review
↓
Deploy
↓
Observability (Product & Tech)
↓
Learn & iterate
Teams, workflows, ways of working optimized for slower coding need to adapt.
What are the opportunities?
-
Product engineering: problem/solutionizing ownership
-
Releasing the review pressure
-
Shipping confidence: tech and product observability
Product egineering
smart image by Andre Barro
Where to start?
-
run/participate in research (qual/quant)
-
develop metrics & monitor the product (Mixpanel)
-
connect with users (interviews)
-
...
-
drive learning and iterating
-
own solutionizing
Context is the new moat
Before: Product -> Engineering context
Now: Product -> Engineering -> AI context
Product and engineering docs (design docs, conventions, ...) matter even more than before. So do codebase patterns
Specialized skills are mindblowing
e.g. /linus-pr-review, /migrate-to-typescript
Reviews
Increased PRs/eng
↓
More reviews/eng
↓
Review fatigue
↓
Quality risk
Paths to further explore
- local skills
- custom, own review GH actions
- trial paid solutions
- first pass with AI to reduce noise
- AI-only review for trivial changes?
Custom GH action
- Orchestrator + summary
- quality review (based on Tomasz's skill)
- conventions review (based on MarkM's GH action)
- <new sub agent - drop-in>
LF contributors and testers!
#eng-gh-reviewer
Observability
Observability
- we ship more frequently, faster
- we'll break things more often
- product will mature
- verbose alerting with clean boundaries
- pattern (tech + product) monitoring
- SLOs
- incidents learnings + CFR monitoring
Action items for tomorrow?
- become friends with Mixpanel and qual research
- AI first pass reviews + trial tooling (internal/external)
- Context and guardrails: docs, specialized skills, ...
- double down on observability
- Do I know how I’ll determine whether it’s a success?
Antrophic paradox
80+% of code at Anthropic is AI-written.
- Boris Cherny, Head of Claude Code, January 2026
We got better at
the obvious bottleneck.
Can we get better at next ones?
AI has moved the bottleneck
By Rafał Rumanek (truti)
