A single build ends when the code is handed off; the FDE loop ends when the customer's metric moves — and the difference decides who owns the outcome.
ConceptWhat it is
A single build runs ticket to code to handoff: the requirement arrives written down, the engineer implements it, and ownership transfers at delivery. The FDE loop runs customer problem to iteration to measured impact: the requirement is discovered, the build is one pass among several, and ownership holds until the outcome is proven. The comparison matters because both shapes use the same engineering skills but distribute risk oppositely — a single build bets that the specification is right; the loop assumes it is not and prices the correction in.
How it worksThe mechanics
In a single build, correctness is judged against the ticket: build what was written, pass review, ship, close. Anything wrong with the specification itself surfaces later, at someone else's desk, as a new ticket. In the FDE loop the specification is a hypothesis: the first pass exists to be validated against the real workflow, 'not enough value' routes back to scoping instead of into a backlog, and the engagement closes on evidence — a metric that moved — rather than on delivery. The loop costs more per use case and buys the certainty a handoff cannot: the person who saw the problem is the person who verifies the fix.
At a glanceSee it
When to use itWhere it fits
- The problem is real but the specification is not — nobody can yet write the ticket that would make a single build correct.
- The cost of a wrong build lands on the customer relationship, not just the backlog.
- AI-shaped work, where behavior on real data cannot be fully specified up front and validation is the only proof.
- Accounts where trust is the product — showing up, iterating and measuring is what earns the next use case.
When NOT to use itLimits & anti-patterns
- Well-understood features inside your own product — a good ticket plus a single build is faster and cheaper.
- High-volume, low-variance work where specification quality is already proven; looping adds ceremony, not certainty.
- When no one will grant the access the loop needs — an FDE without the workflow is a single build with extra meetings.
Trade-offsAdvantages & costs
Advantages
- Impact is verified, not presumed — the engagement cannot close on an unadopted feature.
- Specification errors are caught in days by validation, not months later as churn.
- Ownership never transfers mid-problem, so nothing is lost at the handoff seam.
- Every loop leaves reusable residue: patterns, evals, playbooks.
Trade-offs & costs
- More expensive per use case than a clean single build — discovery, iteration and measurement are all paid time.
- Does not scale by adding engineers the way ticketed work does; each loop wants one owner with range.
- Harder to schedule: 'until the metric moves' resists a fixed end date.
- Applied to a solved problem the loop is pure overhead — the skill is knowing which shape the work in front of you is.
ExampleIn the real world
The clearest tell is what each shape calls done: a platform team closes its ticket when the code merges and review passes, while a forward-deployed engineer at the same company closes an engagement when the customer's queue time, error rate or cost per case has verifiably dropped — the same skills, opposite definitions of finished.
ToolsHow to implement it
- An outcome metric agreed up frontthe loop's definition of done; without it every engagement quietly degrades into a single build.
- Feature flags and staged rolloutmake re-scoping cheap by keeping every pass reversible.
- A re-runnable evalthis loop's validated cases, frozen so the next change cannot silently regress them.
- A decision logsingle builds document the code; the loop documents why each pass was re-scoped, which is the part the product team can reuse.
Cost & effortWhat it takes
Budget the loop at roughly two to three times a comparable single build in engineer time, then weigh that against the cost of a wrong handoff: a single build that misses ships its error into production and into the relationship, while the loop pays for the correction early, on site, where it is cheapest.
What changedWhat changed here
Updated this page Survey figures put agent pilot-to-production failure at 89% and organization-wide scaling at 14%, with the model rarely the bottleneck.
Three kinds of claim, strongest first. Signal runs every morning.