AB-620 deep review — September 28, 2026¶
The AB-620 guide was read in full and mapped to 44 unchanged detailed objectives in nine groups. It now includes six worked examples, ten labs, 48 answered checks and two focused blog exercises. Twenty-four offline assertions check the examples and catalog arithmetic. This is same-context AI research and repair; independent human review is pending. No tenant, model, flow, approval, computer-use session, SDK, load test, telemetry query, deployment or paid lesson was executed.
Exam and learning evidence¶
The official blueprint retains its April 21 page date, no separate skills-effective date and no announced revision. The objective monitor reports no change; accepted snapshot bytes remain unchanged. The credential page lists an active exam, 120 minutes, English and 12 other languages, and no Practice Assessment. The three-day English course is now scheduled for September 30, replacing the guide's September 18 date. That future schedule does not mean the active exam or self-paced paths are delayed.
The three Learn outlines expose 3 + 4 + 4 = 11 modules, still labeled as classic-experience content. Their previous 8h29 combined runtime is historical; the current fetched outlines do not verify it. New study budgets are editorial.
Direct retrieval of both retained Udemy pages was blocked. Indexed public provider pages show Kuljot Singh Bakshi's course updated August, with 12 sections, 64 lectures and 9h27, and Joshua Ravnjak's practice product updated August, with six 60-question tests. The latter's claimed 100-minute real-exam window conflicts with Microsoft's 120 minutes. Originality, difficulty and paid coverage were not audited. Bounded exact-exam searches did not verify a Pluralsight or LinkedIn Learning course; other previously absent catalogs were not exhaustively searched again. Search absence is not proof of market absence.
Objective coverage¶
Every objective has a hashed mapping in the operation record.
| Group | Count | Guide evidence |
|---|---|---|
| Plan an agent solution | 7 | Architecture, audience, identity, channel, governance and reuse |
| Create and monitor agent flows | 6 | Solution/trigger contract, approvals, actions, monitoring, inputs and errors |
| Configure topics | 9 | Tools, flows, formatting, prompts, knowledge, HTTP, generative answers, cards and variables |
| Connect enterprise knowledge | 3 | Copilot/Power Platform connectors, Search and authorization |
| Add tools | 4 | Computer use, MCP, custom connectors and REST |
| Multi-agent collaboration | 5 | Design, Foundry, existing Copilot agents, Fabric and A2A |
| Integrate Azure | 3 | Search/Foundry, model catalog and Application Insights |
| Evaluate performance | 3 | Test sets, methods and result interpretation |
| Implement ALM | 4 | Solutions, existing agents, environment variables and pipelines |
Material changes¶
- Runtime boundaries: correct the stale answer describing the GitHub Copilot harness as wholly preview. Identify individual feature status and prohibit inferred support across harnesses. Scope the current Fabric IQ Data MCP tool to its GitHub harness and distinguish User from Maker credentials.
- Foundry integration: explain the standard-harness preview connector's Activity-protocol prerequisite, its REST/SDK enablement and the portal display limitation. Separate this failure from an older Foundry agent's version error. Distinguish managed prompt catalogs from BYO chat-completions/model restrictions.
- Identity and interaction: separate secure web-channel propagation from authentication publishing, user identity and backend authorization. Explain token-variable changes, background execution with blocked maker credentials, card host versions and repeated submission. HTTP error continuation needs an explicit failure branch.
- Approval and computer use: default AI rejection can bypass a later human stage. Add path arithmetic and a promotion lab for advanced approvals that require recreation after import/sharing. Computer-use supervision can miss a desired pause; enforce required approvals separately and respect intended-use limits.
- Knowledge: document the custom trigger, rewritten query, result contract and shared 15-snippet cap. Add server-side authorization/ranking arithmetic. Keep formal Search connection creation separate from recovery-specific Entra advice; do not recommend hand-built endpoint/key configuration or destructive recovery as a lab. Citations and service authentication do not establish per-document access control.
- Evaluation: replace the old blanket statement that evaluation excludes safety. Current methods include Content Safety and Custom grading, with bounded coverage. Add refusal-specific expectations, test-profile data exposure, explicit denominator metrics, independent latency gates and 89-day export planning. An original confusion matrix exposes missed violations hidden by overall accuracy.
- Operations: distinguish agent-level telemetry from preview environment export, including tables, up-to-24-hour initial delay, unsupported configurations, missing topic/duration fields and nontransactional delivery. Add a peak-pilot extrapolation worksheet; purchased credits do not establish throughput.
Blog intake¶
Jason Moore's September 2 harness roundup supports a feature/harness/status/evidence worksheet. The main article was read; linked whitepapers, videos, tooling and tenant rollout were not verified.
Efrat Gilboa and Dikla Dotan-Cohen's March 26 custom-grader article supports a label-design and calibration exercise. Relevant main sections through custom-label setup were read. The vendor benchmark was not reproduced or adopted as a quality guarantee. Grading measures observed behavior and does not enforce authorization. These are useful supporting readings, not exam authorities.
Evidence and follow-up¶
The operation record is ADLC_Docs/operations/2026-09-28-ab-620-deep-review.json.
It records prior review state, unchanged objective observation, source retrievals,
per-source reading boundaries, objective mappings, blog intake, limitations and
validation results. Broad retained references were refreshed without pretending
that every long article, linked sample or module was re-audited. Blocked direct
Udemy retrieval remains blocked despite indexed catalog observations. The prior
review is retained as history.
The maintenance queue has a September 30 course-availability follow-up and an October 12 integration/evaluation follow-up. The guide, shared learning catalog, source records and public coverage tracker are updated together. Automation detects drift and schedules review; it does not replace manual assessment of support boundaries or independent human review.