The 12 Leverage Points is a two-axis audit map — 4 in-agent levers (Context, Model, Prompt, Tools) and 8 through-agent levers (Standard Out, Types, Architecture, Docs, Tests, Plans, Templates, ADWs) — showing you where engineering investment compounds.

Primitive

12 Leverage Points

The structural split

In-agent resets. Through-agent compounds.

in-agent

The Core Four

Resets every conversation. Whatever you chose for this agent run lives and dies with that session. Per-task calibration.

Context

Model

Prompt

Tools

through-agent

The persistent eight

Persists and compounds. Every future agent run benefits from prior investment. The substrate.

Std Out

Types

Arch

Docs

Tests

Plans

Templates

ADWs

A team focusing on in-agent levers hits a ceiling. A team focusing on through-agent levers compounds indefinitely. The whole framework is built around this asymmetry.

The eight through-agent levers

Each one, its failure mode.

Standard Out

Signal of truth from subprocesses. Fails when progress noise mixes into the result channel.

Types

Named request/response contracts. Fails as untyped dicts — forces caller-by-caller discovery.

Architecture

File/folder structure agents navigate. Fails as deep, inconsistent naming; tests separated from impl.

Documentation

app_docs, CLAUDE.md, AGENT_SYSTEM.md. Fails when stale docs lie about current state.

Tests

The validation harness. Fails when agents can't tell if output is correct without running the app by hand.

Plans

Phased, file-manifested specs. Fails as vague plans that force the builder to re-plan.

Templates

Commands, scenarios, plan skeletons. Fails as copy-pasted prose instead of instantiable skeletons.

ADWs

Pipeline phase scripts. Fails as one-off scripts that don't compose, can't resume, have no state file.

How you know you're improving

Four KPIs. Every leverage point moves at least one.

Size

↑

Size of work handed off per prompt

Attempts

↓

Prompts needed per unit of work

Streak

↑

Consecutive one-shot successes

Presence

↓

How much you must watch/correct

If a KPI is flatlining, the bottleneck is almost always in the through-agent column. In-agent tweaks (better prompt, stronger model) move a KPI within a plateau; through-agent investment moves across plateaus.

What an audit looks like

Score 1–5 per lever. Weakest column = next quarter's work.

app_docs/leverage-points-audit.md — Q2 2026

Context

4 / 5

prime commands exist, CLAUDE.md trimmed last sprint

Model

3 / 5

per-agent choice, but defaults drift to Sonnet

Types

2 / 5

too many untyped dicts in service layer ← focus here

Tests

4 / 5

mirror layout, gates passing, e2e scenarios growing

ADWs

3 / 5

plan/build composites, no test-phase automation yet

Run quarterly or when onboarding. The delta between audits is the measurable progress on the through-agent substrate — the part that compounds.

The discipline

Improvements to through-agent levers compound across every future run.

In-agent tweaks are per-task calibration; they move your KPIs within a plateau. Through-agent investment is the substrate; it moves across plateaus. The audit map surfaces which of the twelve to work on next — and the answer, nine times out of ten, lives in the eight through-agent columns, not the core four.