docs: weekly Copilot Agents Guide update (2026-08-31)

[soyalejolopez]

Executive summary

Weekly audit of the Microsoft Copilot Agents Guide against current first-party Microsoft documentation. This week's pass found material factual errors in the published guide and corrects them.

The most significant: the guide told readers that Cowork local browser use in Microsoft Edge "remains a Frontier capability." It moved to general availability. A FastTrack engineer reading the guide would have told a customer a GA feature was gated behind a limited program.

Four other corrections, each verified against the capability's own documentation page:

#	Was	Now	Why it mattered
1	Cowork GA "worldwide in tier-1 languages"	tier-1 languages, access currently limited to Anthropic-supported regions	"Worldwide" is a scope claim an engineer would repeat in a scoping call
2	Cowork requires "Anthropic approved as a tenant subprocessor"	Anthropic models are optional, enabled separately by an admin	Presented an optional choice as a hard prerequisite
3	Edge local browser "remains a Frontier capability"	generally available, stated together with its admin enablement requirement	Wrong lifecycle state
4	Entra Agent ID opt-out removed "Starting July 2026"	Since May 2026	Wrong date, and wrong tense for a change already in effect
5	Cost: "metered usage in Copilot Chat, tenant-data grounding"	licensed users incur no extra charges for connectors, agents, or plugins; metering applies to unlicensed users accessing shared tenant data	Implied licensed customers face surprise per-use charges

Researcher is additionally rewritten to describe only what its current documentation supports.

---

Public sources consulted

All accessed 2026-08-31. Every source below is public documentation on learn.microsoft.com.

Source	URL	Page date
Copilot Cowork FAQ	https://learn.microsoft.com/en-us/microsoft-365/copilot/cowork/cowork-faq	2026-08-27
What's new in Copilot Cowork	https://learn.microsoft.com/en-us/microsoft-365/copilot/cowork/whats-new	2026-08-27
Use the local browser in Cowork	https://learn.microsoft.com/en-us/microsoft-365/copilot/cowork/cowork-local-browser	2026-08-27
Cowork models	https://learn.microsoft.com/en-us/microsoft-365/copilot/cowork/cowork-models	2026-08-26
Get started with Cowork	https://learn.microsoft.com/en-us/microsoft-365/copilot/cowork/get-started	2026-08-24
Researcher agent FAQ	https://learn.microsoft.com/en-us/microsoft-365/copilot/faq-researcher	2025-10-31
Entra Agent ID in Copilot Studio	https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-use-entra-agent-identities	2026-08-21
Cost considerations for agents	https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/cost-considerations	verified
Manage Copilot Credits	https://learn.microsoft.com/en-us/microsoft-365/copilot/usage-based-billing-manage-copilot-credits	2026-08-11
Work IQ	https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/work-iq/	2026-06-16
Copilot in SharePoint skills	https://learn.microsoft.com/en-us/sharepoint/copilot-in-sharepoint-skills	2026-08-26
FastTrack for Microsoft 365	https://learn.microsoft.com/en-us/microsoft-365/fasttrack/introduction	verified

Checked and confirmed no material delta: Copilot Studio What's New, FastTrack Copilot service description (compared item-for-item), FastTrack introduction, FastTrack Agent 365, Microsoft 365 Copilot extensibility overview.

Key evidence quoted verbatim

• Cowork FAQ: "Cowork isn't an exception, and use and access is currently limited to Anthropic-supported regions."
• Cowork What's New: local browser use "moved from Frontier to general availability."
• Cost considerations: "No extra charges for accessing or using extensibility features (Copilot connectors, agents, plugins)" for licensed users.
• Researcher FAQ: "you can pick the model and mode in the model picker… when in the Researcher Agent. If you are adding Researcher (@researcher) to a Copilot chat outside of the agent, the model picker will be disabled and it will default to auto."
• Researcher FAQ: "it supports over 30 languages" — existing claim re-verified, left unchanged.

---

Files changed

• copilot-agent-strategy/copilot-agents-guide/index.html — 8 Researcher references, Cowork availability card (now with a source link, which it previously lacked), Cowork/structured-agent section intro, Entra Agent ID date and tense, cost guidance.
• copilot-agent-strategy/copilot-agents-guide/README.md — version 3.5.0 → 3.6.0, updated date, Researcher summary, refreshed official links.

---

Validation performed

• tools/catalog-build → node index.js --check → "Catalog metadata is valid for 30 resources." (matches baseline)
• Rendered the guide over local HTTP and walked 21 interactive views with a headless browser: 0 console errors, 0 uncaught page errors. Only pre-existing Tailwind-CDN and in-browser-Babel warnings remain, identical to baseline.
• 14/14 positive assertions for every newly added claim, checked against 57 KB of rendered text rather than source.
• 11/11 negative assertions confirming each corrected claim no longer renders.
• All 6 newly cited URLs return HTTP 200.

The guide is a single-file React app with no build step, so a bad string edit fails silently at runtime. Assertions were therefore run against rendered output, not the file.

---

Two-model review

Two independent high-effort reviewers ran in parallel, then cross-reviewed each other's findings.

• Claude Opus 5 — scope, nuance, confidentiality, editorial
• GPT 5.6 Sol — technical accuracy, completeness, implementation

Both reviewers agreed on the Edge/Frontier error, the Entra date error, and the cost-guidance error. Eight items were contested. Each was settled by re-reading the primary source rather than by preferring a reviewer.

Resolved disagreements

Cowork regional scope — Sol's position adopted. Opus argued the guide's "worldwide" was defensible and that regional limits applied only to Anthropic models. The FAQ forecloses this: "Cowork isn't an exception." Opus had itself proposed the deciding test — that a current-state availability statement should settle it — and the FAQ is exactly that. The competing "worldwide" wording appears only in a June 2026 changelog row, which is weaker authority for present-day scope.

Cost wording — Opus's position adopted, with a correction to Opus's own proposal. Opus was right that "tenant-data grounding" is misleading. But Opus's replacement, "usage-based billing for custom and third-party agents," would have introduced a new error: licensed users are explicitly charged nothing extra for agents. Sol caught this. The committed wording follows the source and neither reviewer's draft.

Edge local browser — both reviewers revised. Adopted Opus's editorial point that GA and the admin-enablement requirement must appear in the same sentence, otherwise a reader concludes the feature is live when the tenant control is still off.

Card length — Opus conceded the constraint was measured wrong (the real ceiling is ~530 visible characters, not 450). Sol conceded its own draft ran long and dropped a specific Edge build number, which would go stale within weeks. The card links to the source instead.

Researcher capabilities — resolved by an evidence check neither reviewer had run. Opus raised the right structural question: if Cowork moved Frontier → GA, did the Researcher capabilities move too? Removing them would then be wrong; they should be relabelled. I checked. There is no current documentation page supporting either a Frontier gate or GA for them, and one of the two names has no presence in Microsoft documentation at all. The only reference is a May 2026 announcement page that also described Cowork as Frontier — demonstrably stale. Removed rather than relabelled, and the guide now describes Researcher only from its current FAQ.

Blog-sourced citation removed. Opus found that the README cited a blog post as the provenance for the Researcher claims being corrected. Fixing the claims while leaving the citation would have invited the same error to return. Removed.

Method finding

The audit method reads feature pages and product-wide release notes, but had no step for per-feature-area "what's new" pages — which is precisely where lifecycle transitions get announced. The Edge/Frontier error survived four prior weekly audits for this reason. Going forward, every lifecycle label in the guide should be checked against that capability's own what's-new page.

---

Unresolved — human judgment needed

From this week

1. Deferred, not rejected: Cowork event-driven tasks, and Copilot in SharePoint skills (Preview). Both are publicly documented and additive. They were held back to keep this PR scoped to correcting wrong claims. Recommend a follow-up PR.
2. The guide still uses "Frontier" in two places for the Microsoft 365 E7 Frontier Suite — a SKU name, not program gating. Correctly left in place. Reviewers should not read this PR as removing Frontier terminology wholesale; at least one model is still documented as Frontier.
3. Two further blog links remain in the README's reference list. They are general announcement links rather than the source of any specific claim, so they were left alone. If the repo wants a strict first-party-docs-only policy for that list, that is an editorial decision.

Carried forward from prior weeks in this rolling PR

4. Harness-scoping consistency across sections.
5. Agent 365 content sits under "Budget & Cost" with no dedicated governance section.
6. fastTrackSupported is modelled per agent type, but FastTrack scope is defined per activity.
7. Agent Builder knowledge limits are not published, so the guide cannot state them.
8. Delta detection against support.microsoft.com remains weak.

---

Confidentiality attestation

Every claim in this PR is sourced exclusively from public Microsoft documentation on learn.microsoft.com, listed above with access dates. No internal, confidential, pre-release, tenant, customer, or personal data appears in the guide, the commit messages, the branch name, or this PR. No secrets are included.

---

Draft PR — not for merge without human review.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[github-actions]

🛰️ PR Sweeper report

Risk: 🟢 LOW · Security gate: ✅ passing · Files: 2

🔒 Automated guardrails (authoritative)

No secret, PII, file-policy, or scope issues detected. ✅

🤖 Dual-model AI review (advisory)

AI review unavailable for this run (models not reachable or no diff). Guardrails above are unaffected.

• GPT-class reviewer (openai/gpt-5): skipped — GitHub Models HTTP 410: {"error":{"code":"github_models_retirement_brownout","message":"GitHub Models is temporarily unavailable as part of a scheduled retirement brownout."}}

• Deep-reasoning reviewer (openai/o3): skipped — GitHub Models HTTP 410: {"error":{"code":"github_models_retirement_brownout","message":"GitHub Models is temporarily unavailable as part of a scheduled retirement brownout."}}

---

The automated guardrails are authoritative and gate the security status. The AI review is advisory and never auto-merges. Thanks for contributing to FastTrack! 🛩️