The most prolific code contributor to Bun's GitHub was Anthropic's AI agent.
Then Anthropic paid millions to acquire the human team anyway.
The code was MIT-licensed. They could have forked it for free.
---
This is the tell.
AI companies publicly say "engineering is over, AI will write all the code."
Then they deploy millions to acquire engineers who already work with AI at full tilt.
That contradiction isn't a mistake. It's a signal.
---
The bottleneck isn't code production. It's judgment.
Anthropic's announcement barely mentioned Bun's existing codebase. They praised the team's ability to rethink the JavaScript toolchain "from first principles."
Translation: we're paying for how these people think. Which tradeoffs they make. Which problems they choose not to solve.
---
AI increases your volume of code. It does almost nothing to increase your supply of people who know which ten lines matter, which PR should never ship, which "clever" optimization will crater your reliability in six months.
When the AI tops the contribution charts and they still buy the humans, that's revealed preference.
---
Leaders don't express true beliefs in blog posts. They express them in hiring plans, acquisition targets, compensation bands.
If you want to know what AI companies actually believe about engineering: follow the cap table, not the keynote.
Read more:
https://t.co/2uGQZ2eVi1
@JasonBud Yeah, I figured. I can see the reasoning behind it - some of the questions Grok asks are insightful and helps create a better solution, hopefully it just needs some tuning.
You are outpacing everyone in velocity of progress though which is what matters long term - keep it up!
Everyone's building "AI agents." 42% of companies abandoned them in 2025, up from 17% the year before.
The problem isn't implementation. It's math.
A 99% accurate agent performing 50 sequential steps succeeds only 60% of the time. Most enterprise workflows require far more than 50 decisions.
We've been treating probabilistic systems as deterministic infrastructure. Expecting software behavior from an extremely sophisticated dice roll.
Software either works or it has a bug.
Models are never fully right and never fully wrong. They are distributions.
Take five AI agents, each "pretty good" at 85% accuracy. Chain them together and your system-level reliability isn't 85%. It collapses to 44%.
That's not a bug. That's probability.
Even at 99% per-step accuracy, a 50-step workflow succeeds only 60% of the time.
Most enterprise processes require thousands of micro-decisions across systems, contexts, and time.
The math does not bend to your roadmap.
Look closely at "successful" production agents. Under the hood you find heavy deterministic scaffolding, hard-coded guardrails, and humans quietly cleaning up the mess.
You're not deploying agents. You're rebuilding traditional software around an unpredictable component and calling it AI-native.
And you pay for it.
Senior engineers writing harnesses instead of features. Domain experts reviewing AI outputs more carefully than they review junior staff. Incident response when the agent does something "no one has ever seen before" (but was always mathematically possible).
Autonomy was supposed to cut headcount. In practice, it drags your most expensive people deeper into the loop.
Stop asking: "How do we make agents reliable enough?"
Start asking: "Where is AI genuinely good, and where is it a liability?"
AI excels at synthesis, drafting, exploration, augmenting judgment. It's terrible at sequential autonomy and anything where 1% errors compound.
Knowing the difference is the skill.
Instead of making the agent the star, use AI to build the show.
An AI coding assistant can help you write actual software. Testable, updatable, predictable, repeatable.
Not as sexy. But it ships. It doesn't delete your production database.
The winners won't have the most agents. They'll have the judgment to know when AI is the tool and when it's the trap.
https://t.co/EVaHGYbfza