Workshop ALDC
Block 03 · Coding Agents
🇪🇸 🇬🇧
Block 03

Coding Agents · architect → spec → conductor flow

The real ALDC flow for development. Three roles that don't overlap. Three subagents orchestrated by the conductor. TDD in seven steps with HITL between phases. No-self-review as a non-negotiable rule.

Maturity levels · where are you today?

Before entering the flow, it's worth setting the bar. Most BC teams have passed through all three levels — but not all in the same order, and many get stuck at the second.

Assisted Copilot

Tab-completion. Spot questions. The dev writes, the agent helps.

Directed Copilot

Structured prompts, prompt files, skills. The dev directs, the agent executes pieces.

Coding agents

The agent orchestrates. The dev validates at gates. Spec-driven + HITL + evidence.

Block 3 is the leap from the second to the third. It's not about using a bigger model — it's about architecture.

Real ALDC flow · 3 moments, 1 workflow between them

al-architect
al-spec.create
al-conductor
Conductor delega en:
al-planning
al-implement
al-review

architect DESIGNS · the workflow DETAILS · conductor EXECUTES with 3 subagents under human supervision

The 3 main roles · they don't overlap

al-architect
Architectural decisions. Identifies applicable skills. Produces architecture.md with ADRs and Skills Applied at the top.
Does not write AL code. Does not generate specs. Does not choose the conductor's model.
al-spec.create · workflow
Expands the architecture into an executable spec — objects, fields, sequences, expected tests. Dialogues until it is closed.
Not an agent. Does not write code. Only produces the spec.md file.
al-conductor
Thin-router. Reads the spec. Delegates phases to subagents (planning → implement → review). Consolidates evidence. Requests HITL at each gate.
Does not generate AL code. Does not invent new decisions. Does not skip gates.
The spec is the brain. The conductor is the nervous system.

LOW vs MEDIUM-HIGH · the flow adapts to complexity

LOW complexity · fast path

For isolated changes, small evolutions, fixes with clear scope.

  • Skips al-architect and the al-spec.create workflow
  • Goes directly to al-developer or the conductor with a minimal plan
  • HITL reduced to a single final gate (PR)

When: <50 lines, no new tables, no external integrations.

MEDIUM-HIGH complexity · full path

For new features, integrations, changes with cross-cutting impact.

  • Full cycle: architectspec.createconductor
  • Plan approval + phase commit + PR · 3 HITL gates minimum
  • Skills Evidencing mandatory in all 3 subagents

When: anything touching more than 2 objects or exposing an API.

The triage is done by the human at the start of the session, or by al-architect if in doubt. The conservative rule is to escalate: when in doubt, full path.

The conductor's 3 subagents

The conductor does not generate code. It coordinates three specialised subagents that do, each with a bounded responsibility.

Planning

al-planning

Reads the spec. Produces phased execution plan. Identifies dependencies between objects. Marks HITL gates. Emits plan.md.

Implementation

al-implement

Executes one plan phase. Writes AL code. Loads relevant skills and declares them in the Phase Summary. Emits phase-N.md.

Review

al-review

Reviews the phase the implementor just wrote. Compliance check against the contract. Skills Compliance table. Emits review in phase-N.review.md.

🚫 Non-negotiable rule · no-self-review
al-implement never reviews what it itself wrote. It is always al-review that reviews, in a separate session with limited context.

TDD in 7 steps · the loop the conductor executes

1 Read spec · extract the phase acceptance criteria conductor
2 Write test first · convert criteria into a failing AL test al-implement
3 Run test · confirm it fails for the right reason al-implement
4 Implement minimum · minimal code to make it pass al-implement
5 Run all tests · ensure nothing existing is broken al-implement
6 Refactor · with tests green, structural cleanup al-implement
7 Review phase · Skills Compliance + HITL gate al-review + human

The loop repeats for each plan phase. At the end of each phase there is a HITL gate: the human approves or requests changes. The conductor does not advance without approval.

Skills Compliance Check · how al-review audits

When closing each phase, al-review produces a table crossing skills declared by the implementor with skills actually applied in the code. If there is a discrepancy, it is flagged.

Declared skillExpected patternEvidence in codeStatus
skill-apiAPI page v2.0 + bound actions/api/v2.0 · POST /incidents/{id}/Microsoft.NAV.resolve✓ PASS
skill-performanceSetLoadFields on repetitive queriesCalcFormula en flowfield, sin SetLoadFields en loops✗ FAIL · Major
skill-permissionsPermission set per new objectCEB Barista Incidents Basic + Full presentes✓ PASS
skill-eventsOnBefore/OnAfter instead of direct modificationPublisher declarado, integration event tipado✓ PASS

The table travels to the conductor's phase-complete.md. If there is a FAIL, the HITL gate is not passed — the human sees the detail and decides whether to accept, request a refactor, or readjust the spec.

Mini-demo · al-architect on Barista Incidents

To close the block, al-architect works from a page and a half of business context — the upcoming Block 4 practical case. Architecture only, no code. The goal is to see the real output and how it maps to skills.

Prompt to the architect

invocation · al-architect on Barista Incidents
Load al-architect.

Context:
Barista Incidents is an extension for a coffee shop chain running BC.
Baristas report incidents (machine breakdown, supply shortage, customer
complaint) from a simple Card page. Incidents have severity levels, are
auto-assigned to a technician queue, and trigger notifications. Once
resolved, they feed a weekly KPI report shown on a Role Center tile.

Task: Produce {barista-incidents}.architecture.md at MEDIUM-HIGH complexity.
Include: ADRs, data model, APIs exposed, events, permissions strategy,
Skills applied at the top.

What should come out of the architect

Replicate in parallel · 3-4 min

Watch how the architect operates on your screen

Anyone with the workspace open can replicate in parallel. The goal is not to produce exactly the same output — it's to see how the architect loads skills on its own and how its own description makes it an invocable agent.

  1. Open .github/agents/al-architect.agent.md before launching. Look for the Responsibilities and Skills Evidencing sections.
  2. Paste the full prompt into Copilot Chat.
  3. Watch how the output starts with > **Skills applied**: — that line is Skills Evidencing in action.
  4. Compare the skills the architect declares with those in our expected list.
What to watch No AL product is touched. The architect reasons in decisions, not code. In Block 4 we take this same architecture and chain al-spec.create + al-conductor to implement it — but that's up to you.

What this block covers

Maturity levels

Assisted Copilot · Directed Copilot · Coding agents. Where the real leap is.

ALDC flow

architect → spec.create → conductor with 3 subagents. Roles that don't overlap.

LOW vs MEDIUM-HIGH

Fast path for isolated changes · full path for anything with impact.

7-step TDD

Conductor loop with HITL at the end of each phase. No-self-review.

Skills Compliance Check

The table that audits declared skills vs those applied in code.

Mini-demo architect

Barista Incidents architecture generated live. Preamble to Block 4.

LOADING README…