Facilitator demonstrates one pattern, participants predict what will happen, solve a related problem themselves, and compare the result against evidence. The goal is to build intuition without requiring model output to match the instructor’s screen.
Arula / Learning design / Lab 2
Agentic Engineering
Learning Architecture.
Stage Overview · One Refund Across a Service Boundary
Canonical high-level model for curriculum design, stakeholder review, and facilitation
Lab 1 teaches engineers to govern an AI-assisted change inside one repository. Lab 2 teaches them to orchestrate multiple scoped AI contexts across a service boundary while the human retains authority over the seam.
01 / Scenario and learning focus
One refund.
A shared responsibility.
A merchant requests a refund. The request arrives in a WSAPI-facing shape, is translated by TTA, and is sent to Payment Processor. For the lab, participants work only on the TTA → Payment Processor portion of the wider refund flow.
- WSAPI
- TTA
- Payment Processor
The real Meridian refund flow continues beyond this seam. CPC, downstream settlement processing, injection, LCS, DCF, the full production routing model, and the complete CI/CD lifecycle are intentionally outside this simulation.
Core Lab 2 questionCan I safely coordinate several AI agents when the engineering truth is split across repositories, contracts, and technical artifacts?
02 / The two learning mechanisms that carry the session
Practice that builds judgment.
The Determinism Ladder
Used as a through-line rather than as the headline of one stage. Each rung removes another decision from the probabilistic model and moves enforcement into a more deterministic mechanism.
03 / Canonical Lab 2 stage map
Seven stages.<br/>Deliberate engineering.
Ground the Work
Lab 2 focus: Frame the Boundary
What happens
Establish the refund scenario, scope, authority, the TTA → Payment Processor seam, and the success criteria before AI is asked to act.
Agentic engineering concept
Context Boundary & Authority
Know what the agent can see, what it cannot assume, and which artifact is authoritative.
Participant
Understand the scenario, shared vocabulary and success checklist; make an initial prediction.
AI
May summarize provided material only.
Human owns
Scope, authority, and what is in/out of scope.
Key output
- Sealed prediction
- Shared vocabulary
- Success checklist
Audit Context
Lab 2 focus: Map the Seam
What happens
Dispatch scoped, read-only agents to inspect each repository independently; reconcile their structured returns into one seam view.
Agentic engineering concept
Scoped Sub-agents & Parallel Delegation
Agents reason locally; the engineer synthesizes globally.
Participant
Write/adapt agent briefs, compare returns, and rule on disagreements.
AI
Inspects one repo or source set in isolation and returns a structured local analysis.
Human owns
Cross-repo reconciliation; any unknown stays unresolved rather than being invented.
Key output
- Repo briefs
- Context / seam ledger
- List of disagreements
Author & Validate the Spec
Lab 2 focus: Make the Spec Buildable
What happens
Use the ledger and technical source material to turn ambiguous requirements into explicit, testable behavior while preserving negative and out-of-scope constraints.
Agentic engineering concept
Spec-as-Context & Readiness Gates
The validated spec becomes the bounded authority agents build from.
Participant
Harden requirements, preserve scope boundaries, and raise open questions.
AI
Helps translate grounded findings into a structured, testable specification.
Human owns
Whether the spec is sufficiently defined and when to escalate rather than assume.
Key output
- Validated spec
- Acceptance criteria
- Open questions
Plan Across Repositories
Lab 2 focus: Design the Orchestration
What happens
Decompose the validated change across repositories; define agent roles, context, tools, expected outputs, dependencies, and a compatibility-safe rollout sequence.
Agentic engineering concept
Agent Orchestration & Context Isolation
Orchestration is deliberate control over who knows, does, and decides what.
Participant
Design the agent workflow and compatibility plan.
AI
May propose decomposition within supplied constraints.
Human owns
Agent boundaries, contract decisions, ownership rulings, and rollout sequence.
Key output
- Orchestration plan
- Repo-scoped agent briefs
- Rollout order + rationale
Build & Validate
Lab 2 focus: Build the Bounded Slice
What happens
Create/approve seam-level tests, then dispatch implementation work within repo-scoped context and explicit engineering boundaries.
Agentic engineering concept
Bounded Agent Execution & Deterministic Guardrails
AI generates inside tests, rules, gates, tool permissions, and contract checks.
Participant
Author/adapt implementation prompts, reconcile returned changes, and keep the pair aligned.
AI
Implements only within its assigned repository and authorized scope.
Human owns
Authorization of changes, contract alignment, and refusal of scope expansion.
Key output
- RED → GREEN tests
- Implementation diffs
- Agent return summaries
Validate with Fresh Context
Lab 2 focus: Prove the Pair
What happens
Run cross-repo/contract checks and use a fresh judging context to compare the combined change against the spec and seam contract.
Agentic engineering concept
Fresh-Context Validation & Independent Judgment
Separate creation from judgment; verify the pair, not just each repo.
Participant
Review findings, verify evidence, and disposition each finding.
AI
Fresh validator judges both repositories without inheriting builder confidence.
Human owns
Finding disposition and final judgment on whether the seam is correct.
Key output
- Contract results
- Validator report
- Finding dispositions
Review, Handoff & Close
Lab 2 focus: Transfer the Learning
What happens
Capture decisions, evidence, unresolved items, and the reusable working pattern that should carry into the next increment.
Agentic engineering concept
Context Handoff, Evidence & Learning Loop
Preserve the minimum durable context the next engineer or agent needs.
Participant
Compare final outcome with the initial prediction and complete the hand-off.
AI
Helps structure evidence and the hand-off.
Human owns
What is accepted, what remains open, and what practice carries forward.
Key output
- Hand-off
- Journey evidence
- Reusable practice
04 / The conceptual progression
Each stage asks
a better question.
- 0
Ground the Work · DEFINE
What is the agent allowed to know and believe?
Lab 2 focus: Frame the Boundary
Agentic concept: Context Boundary & Authority
- 1
Audit Context · DEFINE
How do I split understanding without losing the cross-repo relationship?
Lab 2 focus: Map the Seam
Agentic concept: Scoped Sub-agents & Parallel Delegation
- 2
Author & Validate the Spec · DEFINE
What common technical authority will all agents work from?
Lab 2 focus: Make the Spec Buildable
Agentic concept: Spec-as-Context & Readiness Gates
- 3
Plan Across Repositories · PLAN
Who gets what task, context, tools and responsibility?
Lab 2 focus: Design the Orchestration
Agentic concept: Agent Orchestration & Context Isolation
- 4
Build & Validate · EXECUTE
How do I let AI execute without allowing it to redefine scope?
Lab 2 focus: Build the Bounded Slice
Agentic concept: Bounded Agent Execution & Deterministic Guardrails
- 5
Validate with Fresh Context · JUDGE
Did the combined system actually satisfy the seam?
Lab 2 focus: Prove the Pair
Agentic concept: Fresh-Context Validation & Independent Judgment
- 6
Review, Handoff & Close · JUDGE + LEARN
What evidence and working pattern should survive into the next increment?
Lab 2 focus: Transfer the Learning
Agentic concept: Context Handoff, Evidence & Learning Loop
05 / Design principles for the live lab
Keep the engineering idea
at the center.
- 01
Stage titles communicate the engineering idea. Commands, prompts, demos and checkpoints are implementation details underneath that idea.
- 02
Demonstrate one pattern, then give participants a related problem to solve with their own prompt or agent brief.
- 03
Compare outcomes and evidence, not model wording or screen formatting.
- 04
Keep the human accountable for cross-repository authority, the seam, compatibility decisions, and finding disposition.
- 05
Use deterministic mechanisms for facts that can be mechanically proved; use model judgment where semantic comparison is required.
- 06
Carry Lab 1 concepts forward without reteaching them: fresh context, sub-agents, human gates, deterministic checks, journey and hand-off.
- 07
Do not make hook/skill authorship the conceptual center of Lab 2. Tool authorship can support the exercise, but the core progression is context → orchestration → bounded execution → combined judgment.
06 / What participants should walk away knowing
Carry the practice forward.
Standard lifecycle: DEFINE → DEFINE → DEFINE → PLAN → EXECUTE → JUDGE → JUDGE + LEARN.
- How to identify the context and authority boundaries before using AI.
- How to scope multiple agents to different repositories or evidence sets.
- How to design structured agent briefs and returns so outputs can be reconciled.
- How to turn a distributed, imperfect set of technical artifacts into a build-ready specification.
- How to design an orchestration workflow instead of merely launching more agents.
- How to let AI execute inside deterministic engineering constraints.
- How to prove a cross-repository seam with fresh judgment plus deterministic evidence.
- How to preserve the decisions and evidence that the next engineer or agent needs.
Core principleHumans establish intent, authority and boundaries. AI accelerates bounded execution. Independent evidence supports judgment. Humans remain accountable for the seam and for the final decision. Learning feeds the next cycle.
07 / Document hierarchy
From scenario
to guided practice.
- Scenario Document
- Learning Architecture & Stage Overview This document
- Lab Action Guide
- Facilitator Guide