Nine positions.Argue with any of them.

An unenforced practice is not a practice.

A linter that warns is not a gate. A convention that lives in a senior engineer's head is not a standard. A spec nobody reviews is a document. If it does not block something, it does not exist.

Buying licences is not a strategy.

A team can hand every engineer an AI subscription and watch the bill go up while velocity stays flat. Individual gains do not add up to team gains on their own. What moves is shared context, standard stacks, and honest metrics.

The bottleneck moved. It did not disappear.

Once code is cheap to produce, the constraint becomes trusting it. Doubling coding speed on the 25 to 35% of the lifecycle that coding occupies buys you very little, and only if nothing else is the real constraint.

Write the spec because it makes you think, not because it makes a document.

Most of a spec's value is paid out before any code exists, in the questions it forces you to answer. Agents never ask for clarification. They resolve your ambiguity from training data, confidently, at speed.

The harness is the deliverable.

You do not own the model. You own the rules, skills, tools, tests, and feedback loops around it. That is the part that compounds, survives a model upgrade, and cannot be bought by your competitor.

Context is a budget, not a bucket.

The window is finite and attention across it is uneven. Filling it is not the same as using it. Every MCP server you leave connected spends part of that budget on capability the current task does not need.

Human-in-the-loop is a design principle, not a bottleneck.

Review at the seams, where misalignment is cheapest to catch: the spec, the plan, the diff, the deploy. Calibrate to risk. Reversible work can run hot; payments, auth, migrations and infrastructure cannot.

Resistance is rational.

Engineers who push back are usually protecting something real: craft, ownership, the part of the job they are good at. Answer it with a better definition of the job, not with a mandate.

Your attention is the cap, not your tooling.

Launching ten agents is easy. Holding ten streams of work in your head is not. Find your own number of parallel streams and treat it as a hard limit, because past it the context switching wins.