AI Agent Team for Architecture as Code Initiatives · Architecture as Code: the course

All courses / Architecture as Code / Module 5 / Chapter 20

Learning objectives

After the lesson the participant can:

Describe the multi-agent operating model and the contribution of each specialist role.
Explain how agents are activated by repository events and integrated with automation workflows.
Identify the four categories of decision that must remain with human practitioners.
Define governance and quality measures that make an agent team accountable and auditable.

Key concepts

Multi-agent operating model: The project owner defines priorities and acceptable outcomes; a Project Manager agent turns them into sprint goals and cases and steers information between specialists such as the Architect, Requirements Analyst, Designer, Developer, Quality Control, Editor and Graphic Designer agents.
Living cases: Shared artefacts that accumulate updates whenever two or more specialists exchange information. The focus lies on how communications revise artefacts and decisions, not on tracking activities against a rigid timetable.
Issue-driven agent activation: Agents started by repository events instead of conversation, for example a workflow that runs when an issue receives a certain label, produces a diagram and opens a pull request with a comment on the issue.
Human oversight guardrails: Decisions that stay with people: architectural commitments with long-term consequences, stakeholder communication, security and compliance decisions such as policy exceptions and risk acceptances, and resolution of conflicting recommendations. Agents draft; humans approve.
Agent team quality measures: Lead time from requirement to release within two sprints, automated test coverage of at least eighty-five per cent for critical components, documentation changes within twenty-four hours of a governance or design decision, and fewer than three blockers per sprint.

Interactive exercises

Do them here and get feedback at once. Choose a card and then where it belongs, or drag it.

Interactive exercise 1 · Match

Agent roles and responsibilities

Match each agent role to the responsibility the chapter assigns to it.

Interactive exercise 2 · Sort

What stays with people

Sort each activity by whether the chapter leaves it to an agent as routine work or requires a named human to decide or approve.

Teaching plan

15 min	Teacher presents the operating model and roles, and reads the chapter's note that multi-agent workflows remain an active area of research needing validation and cost monitoring.
10 min	Walk through the collaboration table: for each communication thread, which case is updated and what evidence is left.
10 min	Teacher explains integration through workflows and labels, least-privilege permissions for automation, and the four oversight categories.
20 min	Exercise: write an oversight and escalation policy for an agent team.
5 min	Close on the key insight: agents are most effective as structured participants in a governance framework that keeps humans informed and in control.

Discussion

The chapter treats sprint rituals, daily reporting and quality thresholds as what makes an agent team trustworthy and not as overhead. What would you be unable to audit without them?
Why does the chapter require named human approval, captured in the audit trail, for policy exceptions and risk acceptances?
Which routine activities in your own delivery work could be drafted or checked by an agent, and what validation step would you place after each?

Written exercise

Exercise · 20 minutes

An oversight and escalation policy for an agent team

Assume your own organisation introduces three AI agents into an architecture repository: one that drafts documentation updates, one that checks diagrams and one that proposes decision records. Write the oversight policy. For each agent state what it may draft, flag or execute and which repository event activates it. List the decisions it must escalate, using the chapter's four categories, and name the human role that approves each. Define how conflicting recommendations from two agents are presented to the project owner. Add four quality measures with thresholds, the validation step for output before merge, and how every agent-generated change is traced to a pull request and an accountable individual.

Deliverable: A one-page oversight policy with permitted actions per agent, an escalation table with named approver roles, four quality measures and a traceability rule.

Your work

Self-test

Choose an answer to see whether it is right, or open the answer below each question.

What does the chapter say about external communication such as progress reports to sponsors or responses to regulatory enquiries?
Show the answer
B. Agents draft it and humans approve it before publication
External communication passes through human review before publication. The chapter summarises the rule as agents draft, humans approve.
What coverage do automated tests aim for on critical components in the agent team's quality measures?
Show the answer
D. At least eighty-five per cent
The quality measures set at least eighty-five per cent coverage of critical components, with lead time within two sprints and documentation changes within twenty-four hours of a decision.
When two agents produce conflicting recommendations, what happens in the chapter's model?
Show the answer
A. The Project Manager agent presents both perspectives to the project owner, who resolves it
The Project Manager surfaces the conflict to the project owner with both perspectives clearly articulated, and the human owner resolves it.