Salesforce Agentforce Specialist deep review — September 29, 2026¶
The guide now maps 26 current objectives, answers 40 original prompts and includes 35 executed local checks. The review corrects saved scope wording, strengthens identity/masking/testing boundaries and updates public learning metadata. The workbook is optional practical depth because the official candidate description does not require programming basics.
Canonical wording and maintenance evidence¶
The complete browser-rendered exam article has six groups of 8/2/8/4/2/2 objectives at 20/20/35/10/10/5 percent. The previous saved outline paraphrased those statements and named SOMA where the current objective says generic Multi Agent architecture. Prior objective and status bytes were archived under September 7 paths, and two historical audit references were redirected to the archived objective file. Exact currently observed wording was accepted manually. This is a source-wording correction, not proof of an announced new exam version or effective date.
Direct Help HTML remains a loading/CSS shell, so the automated monitor remains manual review. Browser reading is recorded separately. The Spring ’26 contract and annual maintenance wording were reviewed. The dedicated schedule, readable in English through a German-locale wrapper, lists Summer ’26 availability on August 27, 2026 and a due date of August 20, 2027, in Pacific time and subject to change. The earlier August 24 earned-by cutoff was not reverified in the current read article, unit or schedule. The updated status observation discloses that limit and directs holders to the requirement assigned to their credential; no account was accessed.
Product boundaries and original teaching¶
Masking documentation distinguishes agents and agent-called templates from direct template invocation. The guide avoids treating provider retention language as a claim about all Salesforce/application traces. Service access guidance distinguishes unidentified, verified-contact and supported authenticated-site users, and requires verified-customer query scoping beyond context variables.
Testing Center can modify CRM data and consume requests/credits; proposed evaluation activities therefore use an authorized sandbox and synthetic data. Agent API examples clarify token versus agent identity and clearing provisional chunks after validation failure. Current script documentation distinguishes top-down prompt resolution from model reasoning and one-way subagent transitions from ordinary returning calls.
The three original scenarios combine grounded evidence, a scoped customer action and bounded peer delegation. They require denied-persona tests, trace-to-record reconciliation and critical evaluation outcomes, with no live platform execution claimed.
Actual local execution and its limits¶
The original Python workbook executes 35 checks. An in-memory SQLite fixture filters supplied active/approved/requester labels before ranking. Alice receives A/P2, Bob B/P2 and an anonymous scope P2. Naive global top-two selection yields unapproved X and Bob's B, losing Alice's eligible evidence after filtering. A bound quoted string remains data; deleting P2 removes the anonymous result from this local store.
An exact-fraction evaluator tracks ten expected cases, rejects duplicate/unknown/mistyped results and keeps missing cases in the denominator. Eight routine passes, one authorization failure and one successful handoff yield 90% aggregate but two-thirds across equally weighted categories. Our illustrative gate requires complete coverage, 90% overall and both critical cases passing: that result fails, while a 90% result with only a routine failure passes. These are supplied fixtures, not observed model scores or vendor defaults.
A simplified display state machine clears earlier chunks on validation failure, replaces provisional text with a complete message and rejects invalid local transitions. A capability-set model limits delegated scopes and rejects loops/exhausted hops. No model, Salesforce org/API, paid credit or network call was used; the database closes after use.
Trusted labels do not prove authentication/FLS, and an approval flag does not detect prompt injection. Supplied scores do not implement semantic retrieval. Local deletion does not establish production index/cache propagation. The display is not an SSE parser, and the delegation model is not MCP/A2A authorization. Eight expanded org activities and independent human review remain pending.
Learning catalog and reading boundaries¶
The current Legend trail lists 17h28, replacing the earlier 15h28 observation. Seven timed groups total 1,048 minutes (465/55/155/148/105/105/15), plus an untimed certification step; prerequisites are additional and the path is locked. Public group metadata was read, not all interiors. Summer ’26 maintenance lists five minutes; its main introduction was read without assessment answers.
AFS401 returned a blank body, the Partner Pocket Guide only three extracted characters, and YouTube only title/footer text. O’Reilly and Udemy returned 403. Earlier durations, update dates and claimed coverage are explicitly unverified. The three-column learning table separates current public metadata from earlier observations and our estimates. No paid lesson/practice quality claim is made.
All 27 direct sources have receipts: 25 HTTP successes and two blocked. Several successes are shells or empty bodies, and full current release-hub content was not audited. Browser/indexed fallbacks and selected reading limits are recorded separately. The Trust architecture article was read only in selected identity sections; more specific Service-access documentation supplies channel exceptions.
The operational ledger is ADLC_Docs/operations/2026-09-29-salesforce-agentforce-specialist-deep-review.json. It preserves prior records, hashes, historical archives, manual acceptance, objective mapping, source receipts and reading limits, actual local execution and validation results. The five reserved GitHub guides and disabled notification pilot remain untouched.