Cohort Session 1 Kick off

Session 1 · Kick-off

Why does software with a model in it need a different lifecycle, and which of our projects is one?

Part of the Cohort Kit. Ninety minutes. The session everything else stands on: by the end the room has agreed on what the agentic PDLC changes, has one real candidate project on the table, and has classified it, possibly as not an agent at all.

Pre-reading (send one week ahead; about 30 minutes)

• What is the agentic PDLC? — the four phases, the one hard gate (the sign-off before anything is built) and the eight loops, in one sitting
• How this tutorial works — so the cohort reads the same way
• The evolution of the PDLC — what each earlier lifecycle fixed, and what this one has to
• Why agentic AI projects fail — the six patterns, which the room will recognise

Bring. Every attendee brings one candidate from the team's own backlog that somebody has called "an AI feature". A sentence each is enough. The session picks one.

You will leave with. A shared vocabulary, one candidate project the next seven sessions return to, and an AI-fit verdict on it, written down.

Run-sheet

Minutes	Block	Do this	Material
0–10	Recap	Four questions, one per lesson, each answered by an attendee: what is the hard gate? what do the eight loops close? which earlier lifecycle got closest? which failure pattern have you seen?	—
10–35	The idea	Put the four-phases picture on screen and ask the room what it shows before saying anything. Then the same four phases in the workbench's overview, then one episode of the story so the phases have people in them	Four methods, one lifecycle · Overview · Day 1 episode
35–65	Exercise	Pairs take two candidates from the "bring" list and run each through A1 · Is this AI at all? Then the pairs swap candidates. Compare verdicts before checking against the tree	Exercise A1 · Decision tree 1 · AI-fit assessor
65–80	Decision	The room picks the one candidate the cohort will carry, and records its AI-fit verdict in the product manager's Frame template. If every candidate came out as a rule or a report, that is the finding: pick the least wrong and say why	Templates · product manager · Frame
80–90	Close	Homework named. One sentence each: what changed in how you would start an AI project?	—

Facilitator notes

• Most of a backlog is not AI. Expect A1 to return "rule", "report" or "search" for more than half of what the room brought. Say so before the exercise so it reads as the lesson rather than as a verdict on somebody's idea. The team that learns this in session 1 saves a quarter.
• The room will want to talk about models. Redirect every "which model?" to the phase it belongs in (P1, next session). This session is about whether, not which.
• Ten-minute version if nobody read: the four-phases picture, the hard gate in one sentence ("nothing is built until the spec, the bar and the guardrails exist"), and the six failure patterns read aloud as a checklist against the candidates in the room.
• Two candidates that tie is fine: carry the one with the clearest damage figure, because session 2 needs one.

Homework

The pain register for the chosen candidate: the numbers behind the pain, in the product manager's Discover and Qualify templates. Cases per day, minutes per case, the cost of one wrong case. Estimates are fine; a blank is not. Session 2 derives the bar from these.

Next

Session 2 · Frame and spec — is this an agent at all, what is it worth, and what does a spec it can be built from look like?