Document

Manifesto

The future of coding
is already here

It’s not a dream. It’s a decision.

01 — The reality

AI-assisted coding is no longer a demo toy or a vague promise. Companies are doing exactly that — at industrial scale, on legacy production codebases — today.

Stripe has documented it publicly with end-to-end coding agents (Minions). Experienced developer teams discuss it openly. It’s no longer science fiction. It’s a bold choice. And it’s available now.

The real question isn’t “Does it work?” — the real question is: “Are we ready to change paradigms?”

02 — The paradigm shift

The old model: the developer is the coding machine.

The new model: the developer is the brain directing the machine.

Before After
I write code I define what must be coded
I debug I verify that success criteria are met
I figure out how to do it I decide what to do and why
I’m deep in implementation I’m the architect, analyst, product owner

Humans become 100% focused on:

03 — How it works in practice

Docker as a guaranteed clean room

No more “works on my machine.” The build lives in Docker — a build tuned specifically for Docker that delivers a clean environment every time. docker-compose up means full environment, clean build, tests running.

Full sandboxing so the agent can operate with real operational autonomy: test, break, retry — without risk to the real environment.

The test pyramid as a safety net

For every area of code we touch → we first write characterization tests of existing behavior. The agent doesn’t code in a vacuum; it codes under supervision.

A living documentation base

The agent needs context. That context lives in the repository:

This isn’t documentation for documentation’s sake. It’s the working contract between human and agent. The more precise that contract, the more effective the agent.

04 — The structuring principle: boundaries

A coding agent is only as good as the context you give it. Its context window has real limits — cognitive and physical. This is where the human architect plays an irreplaceable role.

The trap: horizontal coding

Changing serialization strategy across 12 modules, refactoring a cross-cutting pattern across the whole codebase — these are horizontal efforts. They dilute the agent’s context to the breaking point. It loses the thread, produces inconsistencies, misses side effects.

The alternative: bounded vertical slices

Working on a well-defined vertical — payment flow from API to DB, authentication from validation to token — lets the agent hold the whole problem in mind. It operates in a closed world where every piece is visible, dependencies are known, and side effects are containable.

The boundary isn’t a constraint — it’s a success condition. A sovereign agent in a well-bounded world produces coherent, tested code. An agent drowning in a cross-cutting effort produces noise.

This split into integrated verticals isn’t new — it’s what DDD (Domain-Driven Design) calls bounded contexts. What’s new is that these boundaries become the agent’s unit of work.

Defining those boundaries is the architect’s job. It’s hard, intentional, and demands deep domain knowledge. That’s precisely why it’s human work you don’t delegate away.

05 — What changes: the new cycle

The IDE becomes a code review and quality validation tool — navigate fast, critique, question, draft targeted improvement plans. Not a production tool.

  1. Define intent clearly
  2. Document constraints and success criteria
  3. Delegate to the agent + Docker sandboxing
  4. Review produced code (intent vs reality)
  5. Iterate if needed
  6. Validate: functional tests + full stack

06 — Why now?

Because the competitive advantage window is open now. Companies adopting this paradigm compound velocity. Those who wait compound lag.

We already have the ingredients:

What’s missing is the decision and discipline to operate differently.

07 — The future we can build

Every team member becomes a mini product owner for their area of responsibility:

It’s not less work. It’s different work — more strategic, higher impact, closer to the real value we ship.

Potential velocity is unprecedented. Human cost per shipped feature collapses. Quality rises because human brains focus where they add the most value.

08 — Where to start

01
Adopt Docker Clean build every time, tests included
02
Start documentation AGENTS.md, work plans, explicit rules
03
Characterization tests On every area of code you touch
04
Success criteria first Define before coding — not after
05
Trust the agent Focus human energy on review

It’s demanding. Getting started takes real investment.

But the trajectory is clear: every effort invested in this paradigm compounds — docs, tests, and rules accumulate and make the next iterations exponentially faster.

Moving now means choosing sustainable velocity.