Running delivery

The practical core: the twelve steps, bolts and boards, cutting delivery time, reviewing AI-written code, setting and proving the bar, launching safely, cost, guardrails, governance, drift and postmortems.

13 lessonsabout 85 minutesBoards, bolts, launches and the bill

1. 1How to run an agentic AI project: a step-by-step playbookTwelve steps from a vague request to an AI agent in production that someone can defend: who owns each step, what it produces, and what is never delegated.7 min · Beginner
2. 2Bolts vs sprints: planning work when AI writes the codeWhen AI builds a story in hours, a two-week sprint leaves it idle. How to plan in bolts: one unknown each, cut by dependency, integrated the same day.6 min · Intermediate
3. 3How to set up a Kanban board for an agentic AI projectA project board for AI agent delivery: evidence as the exit rule for every column, risk bands as swimlanes, and WIP limits set by review capacity.7 min · Intermediate
4. 4Cut delivery from months to weeks: what shrinks, what won'tAI makes building fast; delivery is more than building. What compresses, what compresses only by design, what never compresses, with the arithmetic for each.7 min · Intermediate
5. 5How to review AI-generated code: by risk, not by diff sizeWhen agents write most of the code, review is the bottleneck. How to band changes by what they touch, route readers by band, and halve a review queue.6 min · Intermediate
6. 6How accurate does an AI agent need to be? The 3-step answerAn AI agent's accuracy bar comes from money, not a round number: divide the damage of a wrong answer by the saving of a right one, per slice.7 min · Intermediate
7. 7Is 82% good enough? How to prove an AI agent meets its barA score from a test set is an estimate with a width. How to report the lower bound, count the cases a slice still owes, and give a verdict a sceptic accepts.6 min · Intermediate
8. 8Shadow mode and canary releases for AI agents, step by stepHow to launch an AI agent safely: run it in shadow beside the people doing the job, cut over at 5% per action, widen on evidence, and rehearse the way back.7 min · Intermediate
9. 9Why your AI agent costs 4× the estimate, and how to fix itAn AI agent's bill rarely has one cause: context, model tier, cache and retries multiply. How to find each in the per-call log and fix them in order.7 min · Intermediate
10. 10AI agent guardrails that hold: why a prompt is not a controlA limit in a prompt can be talked past; a limit in a tool's signature cannot. Six controls that make an AI agent's boundaries real, with a test for each.7 min · Intermediate
11. 11AI governance gates that don't slow delivery: the five gatesFive gates govern an AI agent from idea to wide use: intent, plan, behaviour, release and expansion. Who owns each, and the evidence each one needs.6 min · Intermediate
12. 12AI drift: how to catch the defect with no error messageAI drift is behaviour changing with no deploy and no error. Why accuracy alerts arrive late, what to watch instead, and the two thresholds that catch it.6 min · Intermediate
13. 13Postmortems for AI incidents: find the missing controlAn AI incident postmortem asks which enforced control would have made it impossible, not who was careless. Five moves, from the layer table to the brief.6 min · Intermediate

Also on the wiki.