Cohort Session 7 Govern and Scale - akash-coded/aws-bedrock-agentcore-strands GitHub Wiki

Session 7 · Govern and scale

What holds, what gates, what drifts, and how does an organisation adopt this without rejecting it?

Part of the Cohort Kit. Ninety minutes. The session for the people who will be asked "how do we know it is safe": by the end the room can say where a boundary is enforced and where it is only a prompt, has mapped the gates and who signs, has run a postmortem that names a missing control, and has scored the team's maturity.

Pre-reading (about 40 minutes; the last two are for whoever is rolling this out)

Guardrails that hold — a prompt is a request; a signature is a boundary
The five governance gates — gates that do not slow delivery, and what may halt a phase
Catch AI drift — the defect with no error message
AI incident postmortems — find the missing control, not the string
For the rollout owner: The maturity model · Rolling it out in 90 days · and, if time allows, Team structure and Measuring AI productivity

Bring. The autonomy decision record from session 2. The list of tools or actions the candidate agent can take. Any policy document the organisation already has about AI.

You will leave with. A gate map with the control each tool carries and who signs at each hand-off, an injection test list, a postmortem template the team has used once, and a maturity score.

Run-sheet

Minutes	Block	Do this	Material
0–10	Recap	Where is a boundary enforced? What halts a phase, and what runs alongside it? What does drift look like on a dashboard? What is the difference between a postmortem that names a string and one that names a control?	—
10–35	The idea	The governance page: hard gates and soft gates, then the RACI on the artefacts, so the room sees that signing is by artefact, not by meeting. Then the Day 82 episode: a $2,000 refund that was not owed, and one hour in the postmortem room	Hard gates and soft gates · Who signs · Day 82 · the incident
35–65	Exercise	Pairs do C2 · where a boundary is enforced and G5 · a cap that is not a control on the candidate's own tool list. Then play the incident simulation and run E2: rewrite the postmortem so it names the missing control	C2 · G5 · The incident · E2 · How to hold the security boundary
65–80	Decision	The room maps the control each tool carries and who signs at each hand-off; builds the first three injection tests; then answers the maturity self-check's six questions and records the score and the one thing to fix next	Authority and gate mapper · Injection test builder · Maturity self-check · Gates and governance
80–90	Close	Homework named. One sentence each: which control does the team believe it has that is actually a prompt?	—

Facilitator notes

"It is in the system prompt" is the answer to test. C2 and G5 both resolve to the same finding: a cap the model is asked to respect is not a control; a cap the tool signature refuses is. Have the room find one of each in its own tool list.
The postmortem exercise produces a string the first time. "The model said X" is the symptom. Ask "which line of code refuses?" until the answer is a control that was missing, then write that.
The maturity score will be low, and that is fine. The self-check is by control, not by ambition. Record it; session 8 sets the plan, and the rollout lesson is for whoever owns the ninety days.
Ten-minute version: the hard-gates page, then G5 on the candidate's tools.

Homework

The two-number report, drafted for the candidate as if it had run for a month: the time saved and the money spent, in the workbench's builder, then in the product manager's Learn template. The numbers are projections; the shape is what the sponsor will read every cycle from now on. Two-number report builder.