Manifesto for Agentic Development
We are uncovering better ways of building software with AI agents. Through this work we have come to value:
Declared intent over inferred behavior
Verified properties over reviewed code
System coherence over unit throughput
Named essence over accumulated features
That is, while there is value in the items on the right, we value the items on the left more.
Principles behind the Agentic Development Manifesto
Our highest priority is to deliver software whose correctness can be demonstrated mechanically, by a verifier independent of the generator. The tool that wrote the code cannot be the oracle that confirms it is correct.
Specifications, contracts, and invariants are the coordination protocols for agents. Without them, more agents means more conflict, not more throughput.
Welcome changing requirements — but version the specification before changing it. Agents working against different versions of the truth produce split-brain systems.
The most reliable gate on AI-generated code is not a human reading it but a machine verifying it against declared properties. Human judgment belongs upstream — in deciding what correct means — not downstream in reading what a model produced.
A specification that no one enforces is a wish. Enforcement is mechanical or it is absent. CI is the only reviewer that never skips a check under deadline pressure.
Code is a liability; capability is the asset. Measure properties verified, not lines generated. A codebase that grows without a corresponding growth in verified properties is accumulating unmanaged risk.
The best architectures, requirements, and designs emerge from teams that know which problems are deducible and which are emergent — and refuse to apply the wrong method to either.
Working software is necessary but not sufficient. Working software that satisfies declared invariants is the measure of progress.
Simplicity — the art of maximizing the amount of work not generated — is essential. The near-zero cost of generation makes disciplined subtraction more valuable, not less.
When you can name what your product essentially is, delete what dilutes it — including working code. A product becomes stronger by getting smaller when what remains is the essence.
The best agent-assisted teams maintain two distinct practices: rigorous process for the unit (the spec, the skill, the bounded task) and probe-sense-respond for the system (integration, observation, human judgment at the seams). Neither substitutes for the other.
At regular intervals, the team asks not "how much did we generate" but "how much uncertainty did we retire, and how much of what we built belongs."
Moving human judgment upstream does not mean abandoning the ability to go deep. When mechanical verification fails — and it will — the human must still be capable of understanding the system well enough to diagnose what no gate caught. Judgment that cannot descend into detail is not judgment; it is hope.
The Agile Manifesto was a cure for rigidity — rigid processes, slow delivery, organizations that couldn't adapt. It solved those problems. The Agentic Development Manifesto is a cure for chaos — cheap generation, scarce verification, and systems that grow faster than any human can comprehend. Different era, different disease, different medicine. The problems the Agile Manifesto solved are not the problems we have.
Top comments (0)