Weekly research scan [2026-07-21]

[jgoldfoot]

Draft — for human review. Do not merge without verifying the sources below.

Weekly research scan for the 1–2 weeks to 2026-07-21. Four verified, material findings. All claims are confirmed against primary sources (arXiv abstract pages, official spec changelogs, the benchmark's own site). Anything I could not confirm at a primary source was cut rather than approximated — see Cut / not proposed.

npm test → 122 passed, 7 suites. prettier --check clean. No changes to package.json, package-lock.json, or .github/workflows.

Findings

#	Category	Claim	Primary source
1	[STALE]	MCP revision 2026-07-28 makes the protocol stateless (removes initialize handshake + Mcp-Session-Id) and moves Tasks out of core into the io.modelcontextprotocol/tasks extension. Whitepaper described Tasks as "experimental, added 2025-11-25".	spec changelog, release-candidate post
2	[STALE]	OSWorld 1.0 superseded by OSWorld 2.0 (site notice 2026-06-26). Whitepaper §9.1 claimed CUAs "today complete approximately 45%" and projected toward 100%.	osworld-v1.xlang.ai
3	[NEW]	OSWorld 2.0: 108 long-horizon workflows; best frontier agent (Claude Opus 4.8, max thinking, batched tool calls) completes 20.6% at a 500-step budget, 54.8% partial. Median human time ~1.6h/task; ~318 tool calls vs ~30 in OSWorld 1.0.	arXiv:2606.29537
4	[CITATION]	WebMCP is a Draft Community Group Report (20 July 2026) from the W3C Web Machine Learning CG, exposing document.modelContext. It is not a W3C Standard and not on the Standards Track. Framework described it as MCP "maturing into" browser-native integration.	webmachinelearning.github.io/webmcp
5	[CITATION]	A2A spec is v1.0.0; governed by the Linux Foundation since June 2025, not simply "Google, 2025".	a2a-protocol.org, LF announcement
6	[CITATION]	Odysseys and OpAgent both verified real; added authors + dates. Odysseys: 200 live-web tasks, strongest model 44.5%. OpAgent: 71.6% WebArena, submitted 14 Feb 2026.	arXiv:2604.24964, arXiv:2602.13559

Proposed edits

docs/whitepaper.md

• MCP primitives block: Tasks removed from the core list and moved to a new Extensions subsection alongside MCP Apps. This incidentally fixes a pre-existing mismatch where the text said "three primitives" but listed four.
• New paragraph on protocol-level statelessness and its Layer 5 deployment consequence (sticky sessions no longer required), plus the 12-month deprecation policy and the deprecation of Roots/Sampling/Logging.
• §9.1 retitled "Preparing for 100% CUA Execution" → "The Long-Horizon Gap" and rewritten around the OSWorld 2.0 result. The argument now runs: short-horizon UI competence and long-horizon completion are different problems, and Layers 2–5 reduce the per-step ambiguity that compounds across a long horizon. This strengthens the framework's case rather than weakening it, but it does reverse the section's original "execution is nearly solved" premise — the substantive judgement call in this PR, and the thing most worth your review.
• WebMCP discovery paragraph and the market-signal bullet corrected per finding 4.
• References: OSWorld 2.0 added; MCP/WebMCP/A2A/Odysseys/OpAgent entries updated; list renumbering repaired.

README.md — matching citation updates.

DRIFT.md — one new [QUESTION], see below.

Cut / not proposed

• A current OSWorld 1.0 SOTA number. Secondary aggregators report 85–85.6% and name a model I could not corroborate anywhere ("Claude Mythos 5"). No primary leaderboard confirmed it, so I cut the figure entirely rather than swap one unverified number for a better-sounding one. §9.1 now cites only verified anchors: the 12.24% original baseline against a 72.36% human baseline, and the 20.6% OSWorld 2.0 result.
• WebMCP browser-support claims. Secondary sources contradicted each other (Chrome 146 ships / Chrome 149 origin trial) and all used navigator.modelContext, which the primary spec contradicts — it is document.modelContext. No browser-version claims proposed.
• WCAG 3.0 / ARIA. Checked; WCAG 3.0 remains a Working Draft (March 2026 update) with Recommendation not expected before 2028. Nothing material for the five layers.

Flagged for your decision (DRIFT.md)

Both README and whitepaper recommend <link rel="alternate" type="application/mcp+json" href="/mcp-server" /> for MCP discovery. I could not confirm from any primary source that this media type is registered or that this pattern is specified in either MCP or WebMCP. It may be an original BiModal Design proposal — legitimate to publish, but it currently reads as established practice. Logged as a [QUESTION] in DRIFT.md rather than edited, since rewriting an authored recommendation is your call, not the scan's.

Conventions respected

Benchmark numbers remain directional projections synthesized from agent benchmarks, never presented as a BiModal A/B test. The 20.6% and 44.5% figures are reported as what those benchmarks measured about agents, not about this framework. data-agent-* remains supplementary. npm, not pnpm. Drift in the single DRIFT.md.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 31043ac499

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Keep research questions out of DRIFT

This entry explicitly says the documents agree and only the primary-source status of their shared recommendation is uncertain, so it is not a cross-artifact contradiction. Keeping it under Open turns DRIFT.md into a research/decision log rather than a record of genuine unresolved drift; either fix or relabel the recommendation in the source documents, or track the uncertainty elsewhere.

AGENTS.md reference: AGENTS.md:L252-L255

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Mark deprecated capabilities in the implementation guidance

For readers targeting the newly documented 2026-07-28 MCP revision, this deprecation notice conflicts with the actionable Pattern 11 in docs/implementation-guide.md:957-1015 and examples/mcp-client-capabilities.md:1-75, which still recommend Roots and Sampling without a protocol-version or deprecation warning—they only advise pinning the SDK. Update or version-pin that pattern and example in the same change so current implementers are not directed toward capabilities the whitepaper now says are being retired.

AGENTS.md reference: AGENTS.md:L197-L203

Useful? React with 👍 / 👎.