Recently, Anthropic ran an experiment that should make everyone in tech pause.
They unleashed a swarm of AI agents (parallel Claude instances) and asked them to build a C compiler from scratch.
No daily standups.
No Jira tickets.
No humans writing code.
👉 Result:
~100,000 lines of production-grade compiler code
Builds large systems (even the Linux kernel)
Completed in ~2 weeks
Cost: ~$20,000 in API usage
Yes — the agents actually did the job.
Humans only:
Defined the goal
Set up tests and guardrails
Observed
The agents:
Designed modules
Implemented features
Fixed bugs
Ran tests
Coordinated merges
Iterated autonomously
This isn’t “Copilot helping a developer”.
This is software being built without developers in the loop.
The uncomfortable truth
The bottleneck in software engineering is no longer writing code.
It is:
Defining the problem clearly
Designing constraints
Creating tight feedback loops
Owning correctness and accountability
Code execution has become cheap.
Thinking hasn’t.
My blunt take
If your professional value is:
“I write clean code”
“I implement tickets fast”
“I know this framework really well”
"Its too complex for Agent to work on"
You should be worried.
If your value is:
System design
Architecture trade-offs
Failure modes
Invariants
Business-aligned decision making
You’re entering your most valuable decade.
One strong engineer + an agent swarm
will outperform a 10–15 person traditional team.
This is not the future.
This is already happening.
What part of your job would an agent replace first — and what part absolutely wouldn’t?
First published on LinkedIn.