AB-100 deep review — September 27, 2026¶
The AB-100 guide was read in full and compared with all 74 detailed objectives in ten objective groups. The pass researched implementation gaps, reviewed the original exercises and questions, and applied supported corrections. It adds 11 official sources, associates four existing sources with AB-100, and gives each of the 17 knowledge checks an answer checkpoint. There are now 12 architecture exercises.
This is an AI research and repair pass in the same conversation context, not an independent audit or human review. No tenant deployment, paid course interior, partner-only training, or account-only assessment was tested. The guide's review date describes this documented scope, not a guarantee that every product feature was executed or every recommendation was independently validated.
Exam and credential evidence¶
The public blueprint now displays the October 14, 2026 English outline. Compared with the accepted July 22 snapshot, the differences are the effective-date heading, two additions of “Microsoft” before Foundry Tools, and capitalization of “Service” in one ALM objective. The detailed objective count, groups, and three domain weights are unchanged. The saved July snapshot remains the historical baseline; the October candidate was compared without silently replacing it before its effective date.
The credential page lists the Expert credential and an associate prerequisite. The exam page also lists MB-280 and PL-200, which are absent from the credential page. Both pages contain an unrelated information-protection summary. The guide now warns about these discrepancies, uses the dedicated blueprint for scope, and directs learners to confirm ambiguous prerequisites with Microsoft. This conflict is retained as a blocked freshness finding rather than selecting an unsupported answer.
The public three-day course and 11-module learning path were readable. The path is now titled Architect AI solutions for business productivity. The credential page's empty training widget does not establish that these resources are unavailable.
Coverage and repairs¶
| Objective group, paraphrased | Bullets reviewed | Guide locations and result |
|---|---|---|
| Requirements and grounding | 3 | Parts 1–2: process suitability, data readiness, ownership, and reusable data contracts retained. |
| Enterprise strategy | 13 | Parts 3–5 and 7: CAF, CoE, portfolio, prompts, model customization, cross-platform design; runtime boundaries clarified. |
| Cost and value | 4 | Part 4: added a worked ROI/sensitivity example and a concrete managed-router decision and evaluation process. |
| Agent design | 16 | Parts 2, 5–7: added Foundry capability selection, event-trigger identity, and agent-feed privacy/approval distinctions. |
| Extensibility | 7 | Parts 3, 5–6 and 9: expanded MCP transport/authentication and computer-use identity, supervision, and administration. |
| Prebuilt apps and agents | 7 | Part 7: reviewed workload boundaries, finance-and-operations virtual knowledge and in-app help; retained current-verification limits. |
| Monitoring and tuning | 5 | Part 8: retained layered telemetry, outcome measures, feedback triage, and diagnosis by component. |
| Testing | 5 | Part 8: added native Copilot Studio evaluation constraints and a strategy for generating and independently checking synthetic test cases. |
| ALM | 6 | Part 9: separated standard-harness solution ALM from other runtimes and explained both Foundry publishing models. |
| Security and governance | 8 | Part 10: replaced a mismatched governance citation, retained authorization and threat-model controls, and strengthened exercise checks. |
The per-objective ledger identifies each bullet by position and normalized hash, with a guide-section mapping. It does not reproduce a vendor question bank.
Material implementation findings¶
- Copilot Studio runtimes: The harness overview distinguishes three runtimes. Topic/solution guidance must be scoped to the standard harness; the GitHub Copilot harness name does not establish a GitHub service data boundary. The guide no longer implies one uniform deployment model.
- Autonomous identity: Event-trigger documentation specifies maker credentials and the authentication requirements for autonomous actions. The guide now distinguishes that path from interactive user access.
- Routing: The model-router concepts and implementation guide support explicit eligible-model selection, evaluation, and configuration tracking. A fixed router version name alone is insufficient evidence of a frozen model pool.
- Foundry ALM: The development lifecycle, identity reference, and publishing migration guide describe different resource generations. The guide now differentiates legacy application identities from identities created with new agent objects. Migration requires checking permissions and endpoint consumers; no retirement date was invented.
- Agent feed: The implementation page documents preview status, broad visibility for users with Agent Task table access, and different task semantics. An informational completed task is not evidence of prior approval. The guide and new exercise explicitly reject a confidential per-user approval design based only on this feed.
- Generated pages: Current guidance labels connector and custom-API data paths as preview. Their status is now distinguished from page creation itself.
- MCP and computer use: The MCP connection guide, computer-use configuration, and administrator controls anchor concrete transport, credential, reviewer-access, and disable boundaries.
- Testing and governance: Native agent evaluations have cloud and target limitations. The previous governance citation pointed at a testing page; it is now paired with the actual security and governance overview.
Access and verification limits¶
The live health check covered 73 registered AB-100 source URLs: 68 returned healthy HTTP/metadata responses, five were blocked, and none were missing or errored. The updated guide cites 67 distinct external URLs. HTTP success is weaker than readable evidence: the Whizlabs listing, partner training page, practice-assessment application, and YouTube page did not expose enough content for a new syllabus or runtime verification. Two additional YouTube channel references in the registered catalog also returned minimal shells.
The five blocked requests were the OpenAI supplementary guide, two O'Reilly resources, and two Udemy resources. These were not used to establish new Microsoft behavior. Commercial duration and question/lab counts are explicitly historical observations, not newly verified recommendations. The O'Reilly crash-course and MeasureUp public descriptions were readable; their marketing claims do not establish exam coverage or readiness. Two Microsoft blog duration-signal differences appear in the health report and were not treated as exam or product revisions.
All exercises were reviewed for design coherence, synthetic inputs, access needs, and recovery boundaries. None was executed against a tenant. The ROI arithmetic is a synthetic worked example, not vendor pricing or a business forecast. The answer checkpoints are original explanations of this guide's own questions.
Evidence and follow-through¶
ADLC_Docs/operations/2026-09-27-ab-100-deep-review.json records the guide hashes,
prior review record, blueprint hashes, 74 objective mappings, source-fetch metadata,
health observations, findings, and validation scope. Local source bodies and
candidate differences stay under the ignored
.maintenance/ab-100-deep-review-2026-09-27/ directory; they are not published.
data/source-freshness.json retains the prerequisite conflict as blocked. Recheck
the English blueprint on October 14 before accepting its new baseline, confirm
ambiguous prerequisite eligibility through Microsoft, and execute the relevant
exercises in a disposable licensed environment before claiming hands-on validation.
Blog discovery follow-up¶
A separate September 27 pass searched for useful implementation blogs after the deep review above. Searches combined AB-100 with architecture, Copilot Studio, evaluation, autonomous triggers, and ALM, then followed exact article links from the Microsoft Copilot Studio CAT blog. Named practitioner leads and Microsoft announcement posts were also considered. This was a bounded discovery pass, not an exhaustive blog search or a renewed full objective audit.
The guide's Selected implementation blogs table adds five public articles:
| Accepted article | Author | Published | Updated | Objective fit |
|---|---|---|---|---|
| MCP Servers or Connectors? | Jay Padimiti | 2026-02-13 | 2026-08-02 | Design: integration tradeoffs |
| ALM: The Foundation | James Papadimitriou | 2026-06-03 | 2026-08-02 | Deploy: environments and promotion |
| Quality Gates in Azure DevOps | Adi Leibowitz | 2026-04-19 | 2026-08-02 | Deploy: evaluation automation |
| Copilot Credit Consumption | Petros Feleskouras | 2026-08-25 | 2026-08-26 | Plan/deploy: cost and monitoring |
| Review Before Release | Ramakrishnan Raman | 2026-08-18 | Not separately shown | Deploy: configuration review |
All five are Microsoft-hosted, named-author technical blog articles, registered as Tier 5 expert resources because they are used for examples. Each links to first-party evidence and adds a concrete reading artifact. The guide records runtime, preview, authentication, and sample-support limits beside the resources. Author attribution and source links are retained; each article displayed a CC BY 4.0 notice. No code, illustrations, or substantial excerpts were copied. No dumps, recalled questions, or guaranteed-pass claims were found in these reviewed articles.
Correction retained with the recommendation: The evaluation post's statement about Power Platform pipelines conflicts with Microsoft's newer test-and-deploy guidance, updated August 11. Part 8 now states the documented alternative explicitly. The Agent Review article and current documentation describe different scoring details; the guide recommends a review method without adopting either formula or treating the blog's preview label as a current GA determination.
Other leads were not added:
- Mariano Gomez's ALM account contains useful first-person recovery lessons and cites official documentation. Its author also identifies unsupported record shapes and a single-environment proof of concept. The accepted ALM post is a better introductory reading; this implementation was not reproduced or promoted as production guidance.
- Ivy Fiecas-Borjal's ALM article appeared in discovery but a later open returned 404. Full-content review was incomplete, so it was not added; this does not establish permanent removal.
- Broad launch announcements, existing cross-vendor references, and exam-oriented search results without additional implementation value did not justify more guide links. Suspected dump results were excluded from consideration.
Five blog URLs and six corroborating Microsoft documentation URLs were added to
the source catalog and fetched successfully with readable bodies. Their health
observations are registered, making them part of the existing weekly monitor.
The guide now cites 78 unique URLs: 73 reachable and five access-blocked.
The earlier 67-link guide counts and 73-source live run above describe the deep
review before this follow-up. Original evidence hashes and the blocked
prerequisite finding remain unchanged. This pass does not reset the full
freshness baseline or establish hands-on validation. Local article bodies and
fetch metadata are under .maintenance/ab-100-blogs-2026-09-27/ and are not
published. Acceptance metadata and observations are retained in the source
catalog, source-health baseline, and current AB-100 review record.
Learning-content follow-up¶
The next September 27 pass turned the selected reading into original teaching material inside the guide. A fictional service-case assistant connects five worked examples: consumption per accepted outcome, connector/MCP selection, evaluation completeness, configuration findings, and environment promotion. Each example supplies a decision, failure case, and evidence to inspect. The opening study route links directly to them.
Exercise 13 combines the examples into a paper-based workshop with calculations and an answer guide. Five additional knowledge checks have answer checkpoints, bringing the current guide to 13 exercises and 22 knowledge checks. Earlier counts in this report describe the earlier review phase. The examples, tool names, records, numerical inputs, and predicted results are synthetic; no tenant execution or vendor billing estimate is claimed.
The technical anchors were reread in current Microsoft documentation for ALM, solutions, evaluation automation, connector policies, usage reporting, and Agent Review Tool. Existing registered citations were reused. The guide still cites 78 unique external URLs, and the earlier source-health observations retain their original timestamps. This teaching pass does not renew the full freshness baseline or resolve the official prerequisite discrepancy.
The consumption calculations, incomplete-run percentages, question/answer
alignment, scenario permissions, and internal navigation were checked. An
implementation record with before/after guide hashes and validation results is
stored at ADLC_Docs/operations/2026-09-27-ab-100-learning-content.json.
September 28 scheduled prerequisite recheck¶
The public exam page still includes MB-280 and PL-200, and the credential page still omits both. The existing guide caveat remains necessary. This bounded recheck leaves the blocker open and schedules another check for October 5; it does not reset the September 27 full-review date or establish prerequisite eligibility. Fetch timestamps, response hashes and the extracted lists are recorded in ADLC_Docs/operations/2026-09-28-ab-100-prerequisite-followup.json. No live lab or account-specific eligibility check was performed.