Every practice in software development rests on one assumption: a human will do the work.

True for seventy years.

No longer.

Every practice exists because of a human limit.

Stand-up

mental states must be synchronized

Story points

capacity is finite and variable

Code review

humans make mistakes

User stories

communication is lossy

Roles

one brain can't hold it all

The SDLC is coordination machinery for human execution.

Principles survive. Practices often don't.

Agile is not obsolete. It is incomplete.

There is no AI team.

An AI PM, an AI architect, an AI QA — a human team, just faster?

• Roles exist for limits, coordination, accountability
• An agent is one intelligence, one context
• Job titles import human limits into machines

When you hear an agent described with a job title, ask: what deliverable does it produce?

Negotiable — the most dangerous property

An agent cannot join the conversation. It guesses.

Inference is not negotiation.

• PPrecise — no ambiguity
• RReviewable — at the right level
• EEvidenced — proof of completion
• CComplete — criteria, constraints, failures
• IIndependent — verifiable alone
• SSmall — precise + complete
• EExecutable — machine-actionable

Specifications are the interface

Plain English is not a specification.

Requirements

what was intended

Design

why this approach

Tasks

what was executed

Evidence

did it satisfy intent

Precision is not bureaucracy. It is what makes autonomy governable.

Tokens are not quality.

A million tokens to build the wrong thing is waste — at machine speed.

Correct deliverable → Verified outcome → Business value

Tokens appear nowhere in this chain.

"Which model?" is the wrong first question.

A powerful engine and no steering wheel is not a fast car. It is an accident.

Autonomous coding ≠ autonomous development

Intent → Requirements → Design → Tasks → Implementation → Verification → Deployment → Monitoring
                                              ▲ autonomous coding
                                                addresses exactly one link

• No requirements
• No design
• No evidence trail

Human-in-the-loop is not a checkbox

Both had a human in the loop.

Six levels of human involvement

0 · None

reversible actions

1 · Notification

review after the fact

2 · Approval w/ override

human can block

3 · Approval required

production, compliance

4 · Human-directed

human defines the approach

5 · Human-executed

judgment only

Review intent. Review approach. Review evidence. Code — by exception.

Two lists.

What you built — and what you documented to prove it.

• The requirement — already exists
• The history — already recorded
• The tests — already run
• The approvals — already logged

The evidence package assembles itself.

Eight stages. Four gates. One evidence trail.

Intent
   ↓
Requirements ─────────────── ○ Human review gate
   ↓
Design ─────────────────── ○ Human review gate
   ↓
Tasks ─────────────────── ○ Human review gate
   ↓
Agentic execution
   ↓
Verification ────────────── ○ Human review gate
   ↓
Deployment
   ↓
Evidence (continuous, produced by the process itself)

Not waterfall — cycles in minutes, evidence every time.

The leverage shifts

People don't disappear. Their leverage changes.