~/jamie/lessons/001
LESSON 001 · [DATE] · 4 min · tags: estimating, agents

Estimate in agent-hours, not team-weeks

When the spec is tight, the build is the short part.

The old reflex: someone describes a feature and I hear “a few weeks.” That number comes from human teams — handoffs, standups, context switching, waiting on review.

With a coding agent and a well-defined spec, a working app lands somewhere between thirty minutes and six hours. The bottleneck moved. It's no longer typing; it's how clearly I can describe the thing.

So I estimate in two parts: the time to write a spec I could hand to a stranger, plus agent-hours. If I can't write the spec yet, the honest estimate is “unknown” — not “weeks.”

[ADD A REAL EXAMPLE — e.g. how long the first Captain's Chair build took]

┌─ RULE OF THUMB ──────────────┐
│                              │
│  spec unclear → estimate ?   │
│  spec clear   → agent-hours  │
│  unknowns     → ask first    │
│                              │
└──────────────────────────────┘
15m30m1h2h4h8h1d3d1w2w1mo2moTIME TO WORKING SOFTWARE (log₁₀ minutes)AGENT BUILD · 30m─6hTEAM ESTIMATE · 1w─1moΔ 28× ─ 1440×├ spec-writing time not shown: it lives to the left of every bar
fig. 1 — ranges as stated above, not measurements. [SWAP IN LOGGED BUILD TIMES]

The useful part of the reframe isn't the speed. It's that “how long will it take?” becomes “how well do I understand it?” — and that second question has an answer I can work on today.

[ADD — what you'd tell someone running their first agent build]

$ diff old-habit.txt new-habit.txt
- estimate = gut feel × team velocity
+ estimate = spec time + agent hours