This is very close to what I am trying to formalize with DD-MH.
CLAUDE.md is not just a prompt file.
It is behavioral infrastructure for AI agents.
From the Dualbind perspective, the next layer is:
How do we operate multiple AI agents without collapsing role boundaries, human-held Adoption Right, logs, canonical records, and execution authority?
That is why I published Appendix DD-MH — Dualbind Meta-Harness.
DD-MH is a structural note for operating multiple AI / agent systems under human-held Adoption Right, meaning that final adoption remains with the human.
It treats agents not as magic workers, but as managed participants in a human-supervised workflow:
- roles
- constraints
- verification loops
- logs
- drift control
- human final adoption
- execution boundary design
The key point is not simply “use more agents.”
The key point is designing an operational environment where agents can work in parallel without contaminating canonical records, bypassing human judgment, or drifting into unintended execution.
In that sense, CLAUDE.md is a local behavioral constitution.
DD-MH is my attempt to describe the larger harness structure around this kind of agent workflow.
Version DOI:
doi.org/10.5281/zenodo.20334…
#Dualbind #DDMH #AgentHarness #AIAgent #HumanAI #LLM
Andrej Karpathy didn’t just share a CLAUDE.md file.
He accidentally revealed what software engineering is becoming.
The craziest part?
Most people still think AI coding is about writing smarter prompts.
Meanwhile the best engineers are building entire operating systems around agents.
That’s the real shift.
Karpathy’s workflow philosophy exposed something huge:
LLMs are not powerful because they’re intelligent.
They’re powerful because they can follow systems relentlessly.
And once developers realized that…
everything changed.
That’s why CLAUDE.md files suddenly became a thing across the industry.
Not as prompts.
As behavioral infrastructure.
A way to force discipline onto stochastic models.
Because left alone, AI agents do the same things junior engineers do:
- overengineer simple tasks
- pretend they understand
- rewrite things nobody asked for
- optimize for looking done
- fail silently
So the smartest developers stopped “prompting.”
Instead, they started building environments.
Rules.
Constraints.
Verification loops.
Execution frameworks.
Basically:
turning AI into managed engineering labor.
And now people are running multiple Claude Code agents in parallel like an actual software org:
• one mapping the codebase
• one writing tests
• one debugging failures
• one reviewing diffs
• one researching solutions
• one validating outputs
Not AI assistance.
Agent orchestration.
Karpathy hinted at the most important shift with one idea:
Stop telling the model exactly what to do.
Instead:
define success conditions and let it iterate until it reaches them.
That sounds small.
It’s not.
Because the engineer’s role fundamentally changes.
The bottleneck is no longer:
“who can code the fastest?”
It becomes:
“who can design the best cognitive system around AI agents?”
That’s why this feels different from every AI hype cycle before it.
The leverage is real now.
A single engineer with well-structured agents can suddenly operate with the output of an entire team.
And honestly?
I think most people are still underestimating how massive this shift is about to become.