ROI · Agent-built features

Feature Factory

When does a gated, planned build pay for its token premium?

Features built with agents per month

Feature-shaped work only. Hotfixes and docs go direct.

Share that are wide or backend-critical%

Features whose tasks can run in parallel, or where a bad merge is expensive.

Token multiple on wide features×

Each task gets its own scoped context, gate and verify pass. You pay for that in tokens.

Token multiple on small or serial features×

Serial tasks each re-orient in the same code, so the overhead doesn't parallelise away.

Tokens one agent spends per featuretokens

What a single agent would spend building the same feature directly.

Blended $ per million tokens$

Input and output blended, for the model your builders use.

Features where the gates stop a blocker%

A defect your normal review and CI would have let through. Our own measured case: one foreign-key violation, caught at merge review in 1 of 6 runs.

Cost to fix a blocker before production$

Merged, noticed, reverted the same day, re-fixed by an agent. Usually minutes of human time.

Blockers that would have reached production%

An assumption — set it from your own incident history. Only the blockers that escape are expensive.

Cost of a blocker in production$

Incident time, data repair, customer impact.

Human minutes approving the plan (G1)min

Zero means an agent approved the plan itself and no human reviewed it.

Extra human minutes merging node PRs (G2)min

Scoped PRs to read and merge, beyond the review you would do anyway. This is usually the largest cost.

Loaded cost per engineering hour$

Fully loaded cost of one engineering hour — salary, benefits, overhead. Use your own.

The formula

Nothing on this page is hidden.

premium        = wide share × wide multiple + (1 − wide share) × serial multiple
cost / feature = baseline tokens × (premium − 1) ÷ 1,000,000 × $/M  +  (G1 + G2 minutes) ÷ 60 × rate
blocker cost   = (1 − escape rate) × fix cost + escape rate × production cost
value / month  = features × catch rate × blocker cost

net / month    = value − features × cost / feature
break-even     = cost / feature ÷ blocker cost

Feature Factory spends more tokens than a single agent, every time. Tokens are the small cost. Human minutes at the plan (G1) and merge (G2) gates are the large one, so cutting them moves the result more than anything else.

Most blockers are cheap: caught in review or reverted the same day, an agent fixes them in minutes. Only the few that reach production are expensive, so the calculator weights them by how often that happens.

At realistic values the return is small. If your features rarely produce a blocker, the break-even line tells you to build them directly, and you should.