advisory / 03
Agentic Build Sprint
One workflow, running in four to six weeks — with the evals and the run-book.
Every organisation has one process a person does by hand every week, which a system could carry — and a pilot somewhere that has been at eighty percent for six months. This format is for the first kind, done properly: one workflow, frozen in writing, handover included.
One workflow, four to six weeks, and it is running with an eval suite, a run-book and your team trained to own it.
you receive
You receive: the working system, its eval suite, a run-book, a handover session for your team, and — with your approval — a short public build log with the measured number.
what is fixed — three locked boundaries, one kill gate
- 01
scope frozen to one named artifact, agreed in writing — one assistant, one extraction pipeline, one recurring analysis. Not a programme, not a platform
- 02
the deliverable includes handover — documentation, evals and a run-book, so the system is not dependent on the person who built it
- 03
oversight, not a build shop — no maintenance contract, no hosting, no licence. You own the result and the code
- 04
the eval suite is written before the build is called done: task-level success criteria, and a regression set that survives the next model upgrade — because agent evaluations saturate in months, and a benchmark score is not an acceptance test
- 05
week 1 is a bounded discovery that ends in a decision memo. The sprint can stop there, with the memo, at no penalty — the memo is yours either way
- 06
if the workflow turns out not to be worth automating, that is a finding, not a failure, and it gets written down
engagement shape
4–6 weeks, fixed price per artifact, one workflow. One named internal owner is required: no owner, no sprint — a system handed over to nobody is a system that stops working.
4–6 weeks, fixed price per artifact, one workflow · week 1 ends in a decision memo · stop there, keep the memo
segment variants
- startups
- speed with the receipts: the eval suite and the run-book are what an investor’s technical reviewer reads
- SMB
- the process one person does by hand every week, automated, with a document explaining how it works
- enterprise
- the pilot that has been at eighty percent for six months, taken to done and handed over
- non-tech
- one task, in plain language, with nobody required to read code
- public administration
- a bounded unit description at fixed price, with the documentation the supervisory file expects
not for
not a software house and not a maintenance contract.
the artifacts behind this format
the next step
public administration
a bounded unit description at fixed price, with the documentation the supervisory file expects.
last refreshed 2026-10-17 · reviewed quarterly
Sources: AI Index 2026, Chapter 2 (Stanford HAI).
