Execution is replaceable — the workflows, harnesses, and systems that make agents reliable are not, and that's where every hour of engineering time should land.

Principle

Irreplaceable Work Lives at the Agentic Layer

The split

Two kinds of work run in parallel. Only one compounds.

Application layerreplaceable

db schemas

API routes

React components

unit tests

Leverage: 1×. Every feature you write manually is paid for once — the next one starts from zero.

Agentic layerirreplaceable

CLAUDE.md + context

skills · slash commands

hooks · ADW scripts

triggers · harness

Leverage: N×. Every system you build runs through every future agent invocation — paid forward indefinitely.

The agents can do the left column. They cannot yet build the right column. That's where you live.

The shape of investment

The agentic ring thickens until agents can ship unattended.

Your codebase sits at the center. The ring around it is what you build.

code

Day 1

CLAUDE.md

2 slash cmds

→

code

Month 1

skills · hooks

plan templates

→

code

Quarter 1

ADWs · triggers

full PITER pipeline

Every investment thickens the ring.

When it's thick enough, features ship end-to-end with minimal presence — that is the goal.

A thin ring leaks. A thick ring carries the factory. You build ringward, not featureward.

The time-allocation math

Hours at the execution layer are paid once. Hours at the agentic layer are paid forever.

Same one-hour investment. Radically different payoff curve.

Execution hour

Build one feature by hand

value = 1hr × 1 feature

= 1 unit

× 1

(never again)

Next feature starts from zero. The hour doesn't compound.

Agentic hour

Build the ADW that builds features

value = 1hr × N invocations

= N units

×1

×2

×3

×4

×5

×6

×7

…

Every future agent pays you back on every run, forever.

If the hour doesn't compound, the hour was execution work. Do it twice and it becomes a primitive.

The heuristic

If it happens twice, it's agentic-layer work now.

Anything recurring is a signal. The second occurrence is the promotion.

Explain the same process to Claude twice

manual · in-loop · leaky

→

Promote to a Skill

reusable · cached · cold-startable

Run the same fix by hand twice

manual · re-typed · error-prone

→

Promote to an ADW Script

deterministic · replayable · out-loop

Review the same output shape twice

manual · gated on your attention

→

Promote to a Hook

enforced · silent · blocks at the wall

Second occurrence is the signal. Third occurrence means you missed it.

Where to invest

4 points reset every session. 8 points compound across all of them.

The 12 Leverage Points split on one axis: does the investment persist?

In-agent · resets4

ContextModelPromptTools

Scope: this conversation. Gone when the session ends. Important — but tuning these alone is perpetually re-solving the same task at the same cost.

Through-agent · compounds8

StdoutTypesArchitectureDocsTestsPlansTemplatesADWs

Scope: every future invocation. The substrate. Every hour here makes every future agent faster, cheaper, and more reliable — on every run.

Irreplaceable work lives in the right column.

The engineer who builds systems that build systems outpaces the engineer who tweaks prompts.

To become irreplaceable, you will replace yourself — the version of you that writes each feature by hand.

The discipline

If an agent can do it, you shouldn't be the one doing it — you should be the one building the system that does.

Execution is commodity labor. System design is not. Every hour goes to the ring. Every recurrence becomes a primitive. The engineer who builds the agentic layer keeps their judgment — and replaces the version of themselves that didn't.