Founder directs AI coding while driving Tesla FSD, 5000+ commits per month
CodeRabbit founder argues manual code review is dead; software engineering's bottleneck has shifted from talent to token consumption. Small teams driving AI agents can achieve Facebook-scale release velocity, but the cost is massive compute and CPU bills.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
5.3 Codex made AI agents finally viable for production
Harjit says before July this year they had at most 100-200 PRs per month; in July it suddenly jumped to over a thousand; recently it's exceeded 5000 per month. He points to 5.3 Codex as the inflection point—earlier agents couldn't really get work done, couldn't even deploy directly to production. He says most of his PRs are actually ‘janitorial work’—fixing CI, fixing flaky tests, building caches for distributed systems—not writing features. These are ‘hill climbing’ for the entire software delivery pipeline.
— HarjitSoftware factories let AI agents evolve while unsupervised
Harjit divides AI-driven coding into two modes: one is like clay sculpting, refining the UI message by message manually; the other is ‘factory’ mode, where you give the agent a goal—say, migrating git commit logs to object storage—and let it run for six weeks climbing the hill on its own. During that time he didn't read a single line of code. He argues that the deployment infrastructure that used to only companies like Facebook or Netflix with tens of thousands of people could afford—canary deployments, observability, CI/CD—now even 30-40 person teams with agents must have, or they can't scale their release velocity.
— HarjitOnly engineers with clean records earn direct production access
CodeRabbit has set up something like a ‘driver's license’ mechanism for engineers: if someone's code has unfixed issues lurking in production for a long time—say, a regression discovered 100 hours after deployment—they lose their spot on the ‘review-exempt’ list and must pass through manual code review for a period before regaining direct deploy privileges. This ties accountability and trust to individual history, not applying uniform review to everyone.
— HarjitLet code chaos happen first, then deploy cleanup agents
Harjit's approach is ‘let chaos happen first, then clean it up’: let the agent generate code at speed, tolerate code bloat, weaker security posture, and test redundancy in the near term, then run a dedicated batch of ‘test dehydration’ agents to continuously prune flaky tests, eliminate redundancy, and refactor abstractions. This is more realistic than preventing entropy upfront because once code volume explodes, you don't have time to design everything right from the start.
— HarjitCPU is now the scarcer bottleneck than GPU
Because agents run CLI commands and massive CI and simulation testing, CodeRabbit's CI bill was quoted at $13 million per year, so they built their own. He says even GCP has very limited available compute—‘Anthropic took it all’—and they couldn't even upgrade database instances in central regions due to lack of compute. He believes models themselves are now commoditized; the real bottleneck is shifting to data centers and energy.
— HarjitEight-million-dollar seed rounds no longer fund software factories
Harjit said on Twitter that early-stage startups now need ‘software factory’-grade quality tokens and massive compute to run CI, sandboxes, and infrastructure, implying the old model of a few smart engineers plus an $8 million seed round is becoming obsolete. He argues the bottleneck has shifted from ‘paying for talent’ to ‘driving token consumption’—one person can now drive $2-3 million in annual token spend. The host pushes back, arguing this is specific to the AI DevTools space, not a universal startup dynamic.
— HarjitThe competitive window has collapsed from months to weeks
Harjit says a good idea plus solid product work used to buy you six to nine months of competitive moat, but now ‘even Anthropic will copy you in two weeks’. He argues this is forcing VC-backed startups to burn more tokens in ever shorter windows for speed, because the window to capture user mind share has compressed to weeks, not the months of the past.
— HarjitWaking at 3am for agents is not polyphasic sleep optimization
Harjit admits he's changed his sleep cycle, waking at 3am to issue a few commands to stuck agents before sleeping again, comparing this to pre-industrial farmers tending livestock at night and calling it ‘polyphasic sleep’. The host explicitly disagreed at the end, arguing this is sacrificing sleep for the obsession with keeping agents ‘running 24/7’, which is not efficient. He suggested that if he really wants agents running round the clock, he should put them in the cloud instead of sacrificing his own sleep.
— Host/HarjitIn their own words · checked verbatim
I mean, the human code review is dead, but the human review is not dead. Some review still has to happen.
Harjit25:38
And once you have like enough strikes, we take your driving license away. We make you go through the code review.
Harjit30:44
I mean, this might be a hot take, but like humans were writing slop, not AI. I mean, that's what I'm realizing.
Harjit44:53
So our AI cost, I mean, mine alone would be $250,000 a month, $275,000, something like that.
Harjit51:59
So I think it's more like a polyphasic sleep. Maybe I'm like... Exaggerating the health.
Harjit1:13:27
He calls this polymorphic sleep. But to me, it sounds like deprioritizing sleep all to unblock agents. This feels very inefficient to me.
Host1:24:40
Figures
| CodeRabbit monthly PR count | over 5000 | 0:00 |
| CodeRabbit team size | approximately 40 engineers | 18:30 |
| Original CI vendor annual quote | $13 million | 39:51 |
Glossary
- Hill climbing
- An AI agent iterating within a self-validating loop to continuously improve a system.
- Software factory
- Development model where AI agents replace humans for most coding, testing, deployment, and operations.
- Polyphasic sleep
- Sleep divided into multiple segments, waking at night to handle tasks before resuming.
How to listen
DevTools and AI coding tool founders; engineering leaders tracking AI's impact on development workflows and organizational structure; technical managers interested in a high-intensity AI usage case study.