Tech Radar: PostgreSQL 19 GA Week, KV-Cache Hardening Checklist, AAIF at 10 Projects

Answer-first: This edition covers the week the data-platform calendar and the AI-infrastructure incident ledger converge. Data platforms: PostgreSQL 19 GA is scheduled for November 19 (per the official RC2 announcement), arriving on the heels of the November 12 drop that fixed 52 security vulnerabilities across 18.7/17.12/16.16/15.20/14.25 and retired PostgreSQL 14 — we turn the incident details into a concrete multi-tenant KV-cache and database hardening checklist. AI infrastructure: the Cloudflare GLM-5.3 KV-cache post-mortem (155 affected responses, all prompts over 681,280 tokens) and vLLM’s Mooncake transfer-id collision (CVE-2026-94627, CVSS 8.7) now have published fixes — SGLang v0.5.22 and vLLM v0.32.0 — that convert directly into an inference hardening checklist. Agent protocols: the Linux Foundation’s Agentic AI Foundation is at 10 hosted projects after adding Agent Skills, Open Responses, OpenEnv and OpenSandbox, with the MCP Dev Summit in Nairobi scheduled Nov 19-20. Toolchain: Go’s eMudhra aftermath formalizes same-day rebuild discipline with govulncheck -mode=binary. Plus quick hits: Cloudflare changelog (AI Search + D1, Hyperdrive limits), GPT-6 Cyber, Claude Fable/Mythos 5.2, and the AWS re:Invent preview.

Prerequisite: Readers should start from last week’s digest — Tech Radar: KV-Cache Isolation, Go 1.27.3 Urgent, A2A — plus the deep dives on agentic memory architecture, MCP stateless HTTP & A2A v1 wire protocol, and vLLM v1 production KV-cache optimization. The PostgreSQL 19 companion deep-dive is PostgreSQL 19 GA — The Production Upgrade Guide.


name: "Weekly Radar 2026-W47: PostgreSQL 19 GA Week, KV-Cache Version Floors & AAIF Consolidation"
ring: "Adopt"
quadrant: "AI Infrastructure & Platform Engineering"
rationale: "Replaces 'patch when convenient' with two dated obligations this week — the November 12 PostgreSQL security drop (52 CVEs, PG14 EOL) before the November 19 GA evaluation, and pinned inference-engine floors (SGLang >= 0.5.22, vLLM >= 0.32.0) with checksum/canary verification — while standardizing on AAIF-governed A2A v1.0.0 + MCP 2026-07-28 and same-day Go rebuild automation."
adr_link: "/radar/2026-11/radar-2026-11-19-weekly-digest-kv-cache-a2a-go-hardening/"
justification: "The Cloudflare post-mortem proves KV-cache pool/config mismatches leak cross-tenant context without raising an HTTP error (155 responses, all prompts over 681,280 tokens), vLLM CVE-2026-94627 shows the same failure class on disaggregated KV transfer, and the PostgreSQL November drop ships 8.8-rated code-execution CVEs in backup tooling — three independent evidence chains all pointing at versioned floors plus verification gates rather than prompt-level or config-level mitigations."

1. The Week at a Glance: Radar Ring Matrix

RingTechnology / PatternDomainWhy This Week
ADOPTNovember 12 PostgreSQL point releases + post-update reindex sweepData Platforms52 security vulnerabilities (several 8.8, code execution via pg_restore/pg_dump); the reindex/re-backup steps are part of “patched”, not optional extras.
ADOPTInference-engine version floors with verification flagsAI InfrastructureSGLang v0.5.22 carries the Cloudflare-class fixes; enable --disaggregation-enable-kv-checksum and KV canary detection where supported.
ADOPTSame-day Go rebuild discipline (go version -m + govulncheck -mode=binary)Languages & FrameworksThe eMudhra aftermath formalized the URGENT track; trust-store incidents bypass source-level CVE scanners entirely.
TRIALPostgreSQL 19 in staging (GA scheduled Nov 19)Data PlatformsREPACK (CONCURRENTLY), parallel autovacuum, AIO auto-scaling, WAIT — see the companion deep-dive for the full upgrade playbook.
TRIALAAIF-governed dual-stack agent wire (A2A v1.0.0 + MCP 2026-07-28)AI & Agentic SystemsAAIF now hosts 10 projects; MCP Dev Summit Nairobi Nov 19-20 keeps the stateless-core cadence moving.
ASSESSCloudflare AI Search + D1 data sources (beta)Cloud NativeIndex D1 rows as search documents directly — collapse a vector-ingestion sidecar for small-to-medium stores.
ASSESSGPT-6 Cyber (vetted access)AI & Agentic Systems“Critical” rating under OpenAI’s own Preparedness Framework, 48.5% benchmark score, 99.93% reliability on reported tests — assess security implications before adoption.
HOLDPostgreSQL 14 everywhereData Platforms14.25 was the final release; the branch is EOL as of November 12, 2026.
HOLDvLLM ≤ 0.29.0 with Mooncake / SGLang < 0.5.21 for multi-tenant servingAI InfrastructureCVE-2026-94627 (CVSS 8.7) and the Cloudflare class both live below the version floors.
flowchart TD
    subgraph ADOPT["ADOPT (Apply This Week)"]
        A1["Nov 12 PG point releases<br/>+ reindex / re-backup sweep"]
        A2["Engine floors: SGLang >= 0.5.22<br/>vLLM >= 0.32.0 + KV checksum/canary"]
        A3["Same-day Go rebuild:<br/>go version -m + govulncheck -mode=binary"]
    end
    subgraph TRIAL["TRIAL (Controlled Trials)"]
        T1["PostgreSQL 19 staging<br/>(GA scheduled Nov 19)"]
        T2["A2A v1.0.0 + MCP 2026-07-28<br/>dual-stack under AAIF"]
    end
    subgraph ASSESS["ASSESS (Evaluate)"]
        S1["Cloudflare AI Search + D1 (beta)"]
        S2["GPT-6 Cyber (vetted access)"]
    end
    subgraph HOLD["HOLD (Stop or Avoid)"]
        H1["PostgreSQL 14 (EOL Nov 12)"]
        H2["vLLM <= 0.29.0 + Mooncake<br/>SGLang < 0.5.21 multi-tenant"]
    end
    ADOPT --> TRIAL --> ASSESS --> HOLD

2. Theme 1 — PostgreSQL 19 GA Week: The Scheduled Date, the Security Drop, and the EOL Cliff

BLUF: November 19 is the scheduled GA date for PostgreSQL 19 per the official RC2 announcement, but the urgent item already shipped on November 12: 18.7/17.12/16.16/15.20/14.25 fixed 52 security vulnerabilities and over 140 bugs, PostgreSQL 14 reached end-of-life with 14.25 as its final release, and several fixes require post-update reindexes or a fresh full backup. GA day is an evaluation day, not a production-upgrade day.

The November 12 announcement is the week’s most consequential document for data-platform teams:

  • 52 CVEs fixed, including CVE-2026-6460 (8.8 — type confusion after ALTER ... SET DATA TYPE executing arbitrary code), CVE-2026-116314 (8.8 — pg_restore arbitrary code execution), CVE-2026-116318/116330 (8.8 — SQL/psql command injection in pg_dump output), and CVE-2026-120145 (8.8 — logical replication writing to subscriber system catalogs).
  • PostgreSQL 14 is end-of-life. Per the versioning policy, the branch will never receive another security fix.
  • Post-update steps are mandatory: reindex affected GIN/pg_trgm GiST/TOAST indexes, reindex lengthless-bpchar B-tree indexes, take a new full backup if you use incremental backups, and clean up invalid NOT NULL NO INHERIT constraints on partitioned tables.
  • PostgreSQL 19 RC 2 shipped in the same drop, documented as “mostly identical to the final release,” with the revert of JSON_ARRAY()/JSON_OBJECT() from expression indexes, generated columns and partition keys — drop and redefine any such objects created during beta/RC testing.

The RC1 announcement confirms the schedule context: RC1 shipped October 15, RC2 on November 12, GA planned for November 19 — roughly six weeks past the traditional cadence, bought in exchange for a dense RC series. Headline features for the GA evaluation: in-core REPACK (CONCURRENTLY), parallel autovacuum with scoring-based prioritization, auto-scaling async I/O workers (io_min_workers/io_max_workers), the WAIT command for standby read-your-writes, and logical replication that no longer requires a restart when flipping wal_level = replica.

Radar Take: (Opinion, clearly separated from the facts above.) Treat the week as a three-step ladder: patch the running majors first (the 52-CVE drop is live risk), retire PG14 next (the clock has stopped), and evaluate 19 in staging last (full playbook in our companion deep-dive, PostgreSQL 19 GA — The Production Upgrade Guide). Gate production 19 on 19.1 — GA-plus-one is the standing rule for data platforms. For the service-side patterns that depend on Postgres transactional semantics, the Temporal saga patterns work is the place where stale-replica reads were first flagged as a correctness risk — 19’s WAIT command is the database-layer answer.


3. Theme 2 — KV-Cache Hardening: The Checklist the Incidents Already Wrote

BLUF: Two incidents now have complete published fixes, and together they define the multi-tenant inference hardening checklist. Cloudflare’s post-mortem shows a config/pool mismatch (1,048,576-token context over a 681,280-token prefill KV pool) triggered an SGLang bug that let decode attend over stale pages — 155 affected responses, all prompts longer than 681,280 tokens. vLLM ≤ 0.29.0’s Mooncake connector leaks GPU KV-cache blocks via transfer-id collision (CVE-2026-94627, CVSS 8.7). The fix floors are SGLang v0.5.22 and vLLM v0.32.0 — plus verification flags, not just versions.

3.1 The Cloudflare Post-Mortem, In Numbers

The Cloudflare remediation post is the single most detailed multi-tenant KV-cache incident write-up published to date. The verified mechanics:

  • Window: August 27 – September 9, 2026; 155 affected responses, all to prompts longer than 681,280 tokens. Reported September 8 by researcher Gadi Evron (Knostic) via HackerOne; the model went offline within 40 minutes of incident declaration and was restored September 9.
  • Root cause was two overlapping failures. A deployment misconfiguration: GLM-5.3 ran a 1,048,576-token context window over a prefill KV pool sized for only 681,280 tokens — and testing validated HTTP codes and streaming, not output coherence, so the mismatch never surfaced. Then an SGLang bug: an over-length prompt was internally rejected and replaced with a one-token placeholder, but disaggregation admission queues checked whether a request was finished, not whether it was rejected — so the placeholder prefill ran, one page of KV was transferred, decode assumed the full cache was overwritten, and the response was generated attending over stale cache pages of earlier requests, returned as a clean HTTP 200.
  • The fix pair: rebalance GPU memory so the prefill pool holds about 2.04 million tokens (~2x context length) — valid requests never hit the rejection path again — and patch SGLang so rejected requests are retired at the admission queue before any KV transfer (proposed upstream as SGLang #38704; related fixes in #36988/v0.5.20 and #38935/v0.5.21). SGLang v0.5.22 includes the fixes.
  • Hardening added: deferred KV release extended to all failure types, bounded Mooncake transfer batches, load balancer no longer aborts mid-transfer on client disconnect, and a deployment gate that checks KV pool size against the running engine’s context length. Operators are advised to enable --disaggregation-enable-kv-checksum and KV canary detection where supported.
  • The stated principle: KV-cache pages are not zeroed between requests (for performance), so page ownership tracking is a security boundary — treat lifecycle bugs as tenant-isolation bugs, with the long-term goal of page-level tagging checked at every read (“integrity by construction”).

3.2 vLLM: CVE-2026-94627 and the 0.32.0 Security Drop

The VulnCheck advisory (CVE-2026-94627, CVSS 8.7, discovered September 21 by Mingkai Yu, Jiapeng Li, and Jiajia Liu) documents the same failure class on vLLM’s disaggregated KV transfer: the Mooncake connector fails to track GPU KV-cache block ownership when concurrent child requests share a single transfer ID (CWE-401), letting orphaned KV blocks accumulate until process restart — high availability impact, and the same page-ownership boundary failure Cloudflare described. The fix landed in vLLM PR #49796 and the surrounding hardening wave.

The vLLM release stream shows what a security-hardening release looks like from inside: v0.32.0 (Oct 18) fixed three advisories in one drop — chat-template resource exhaustion (GHSA-4hhp-h66f-j5j7), tokenizer max_token_id (GHSA-q97c-p368-75c4), and P2P ZMQ socket exhaustion (GHSA-frvg-qw88-hww9) — plus hardened CORS defaults, raster-only untrusted images, signed shared-memory handles, and deep-nested JSON schema rejection. v0.31.0 (Oct 5) added prefix-cache key tagging so a LoRA name and cache_salt cannot collide. v0.33.0 (Nov 2) continues the hardening pace (PyTorch 2.14, /v1/decisions endpoint, CUDA 13.0-only artifacts).

3.3 The Multi-Tenant Inference Hardening Checklist

Combining both post-mortems with our vLLM KV-cache optimization radar, the checklist a platform team should implement before the next engine upgrade:

# 1) Pin versioned engine floors — fail CI on drift
#    SGLang >= 0.5.22 (carries Cloudflare-class fixes)
#    vLLM   >= 0.32.0 (three GHSA fixes + hardened defaults)
sglang --version   # must be >= 0.5.22
vllm --version     # must be >= 0.32.0

# 2) Enable transport-level integrity where supported
#    (per Cloudflare's guidance for disaggregated prefill/decode)
sglang serve ... --disaggregation-enable-kv-checksum
#    + enable KV canary detection where your engine supports it

# 3) Gate deployment config against the RUNNING engine, not the YAML
#    Cloudflare's own remediation: KV pool must be sized >= ~2x context length
#    so the rejection path is never reachable by valid traffic.
#    (Automate: query engine-reported context length + pool size at boot;
#     abort deployment on mismatch.)

# 4) Two-session canary isolation test in serving CI:
#    session A writes a canary token; session B's output must never contain it.
#    If the stack cannot pass this, you do not have multi-tenancy.

# 5) Binary-level scan of the shipped engine artifact
govulncheck -mode=binary ./bin/inference-gateway   # Go control plane
#    (for Python engines: pin via lockfile + image digest, scan the image SBOM)

Radar Take: (Opinion.) The two post-mortems converge on one architecture principle: cache reuse is an economics feature; cache isolation is a security boundary — and the second cannot be derived from the first. Cloudflare’s “measure pool vs context length directly on every node” and the checksum/canary flags are the operational form of that principle. Our agentic memory radar recommends KV-cache-aligned memory replay for token economics — that advice stands, with the added hard constraint that cache affinity never crosses tenants.


4. Theme 3 — Agent Protocols: AAIF at 10 Projects, MCP Stateless Core, Nairobi Next

BLUF: No new spec revision landed in the Nov 13-15 window — and that stability is itself the news. The Agentic AI Foundation now hosts 10 projects after the October 22 addition of Agent Skills (Anthropic), Open Responses (OpenAI), OpenEnv (Hugging Face) and OpenSandbox (Alibaba); A2A remains stable at v1.0.0 with 150+ organizations; MCP’s latest revision remains 2026-07-28 (stateless core). The next calendar anchor is the MCP Dev Summit Nairobi, November 19-20.

Per the Linux Foundation press release (AGNTCon + MCPCon North America, San Jose, October 22-23):

  • AAIF grew from 3 launch projects (December 2025) to 10 hosted projects in its first year: Agent2Agent (A2A), Agent Router, agentgateway, AGENTS.md, goose, MCP — plus the four new additions: Agent Skills (reusable capability packages for agents, contributed by Anthropic), Open Responses (multi-provider LLM interface spec based on the OpenAI Responses API, contributed by OpenAI), OpenEnv (containerized RL environments exposing MCP and Gym interfaces, contributed by Hugging Face), and OpenSandbox (sandbox platform for AI apps with Docker/Kubernetes runtimes, contributed by Alibaba).
  • A2A governance and adoption: Google transferred A2A to the AAIF in August 2026; the April 2026 milestone press release documented 150+ organizations, embedding in Azure AI Foundry/Copilot Studio and Amazon Bedrock AgentCore, with Signed Agent Cards for cryptographic agent identity in v1.0.
  • MCP: the 2026-07-28 specification remains the current revision — stateless core, no Mcp-Session header, tasks as an official extension — and Google’s adoption write-up describes scaling agent infrastructure on plain HTTP because of it. Our Nov 7 radar covers the wire-level design.
flowchart TD
    LF["Linux Foundation<br/>Agentic AI Foundation<br/>10 hosted projects (Oct 22)"] --> CORE["A2A v1.0.0 (150+ orgs)<br/>+ MCP 2026-07-28 (stateless core)"]
    LF --> NEW["Oct 22 additions:<br/>Agent Skills (Anthropic)<br/>Open Responses (OpenAI)<br/>OpenEnv (Hugging Face)<br/>OpenSandbox (Alibaba)"]
    LF --> INFRA["Agent Router + agentgateway<br/>+ AGENTS.md + goose"]
    CORE --> SW["Dual-stack agent platforms:<br/>A2A for delegation,<br/>MCP for tool calls"]
    NEW --> SW
    INFRA --> SW
    SW --> NEXT["Next anchor:<br/>MCP Dev Summit Nairobi<br/>Nov 19-20, 2026"]

Radar Take: (Opinion.) A week with zero protocol churn is the strongest evidence yet that the interop layer has stabilized: procurement can pin A2A v1.0.0 + MCP 2026-07-28 without fearing a mid-quarter break. The four new AAIF projects are worth one tracking line each — Agent Skills in particular is the capability-packaging layer that our agentic memory architecture analysis assumed would eventually be standardized. Watch Nairobi next week for the stateless-core adoption cadence.


5. Theme 4 — Go: Same-Day Rebuild Discipline Becomes Standard Practice

BLUF: The eMudhra aftermath (Go 1.27.3/1.26.10, released under 7.5 hours from notification on October 19 — the first URGENT-track release in Go’s history) is now the reference playbook for toolchain incidents. Two durable consequences: the Go team committed to Linux revocation-check parity (tracking the upki project in issue #82402), and platform teams formalized same-day rebuild pipelines validated by go version -m and govulncheck -mode=binary.

The Go blog’s incident write-up details the timeline worth internalizing: OneCRL blocked the eight eMudhra intermediates October 17; researchers demonstrated the keys were trivially weak (recoverable “in microseconds”) on October 19; the Go security team was alerted at 11:45 PDT, finalized the patch at 14:00, began release at 15:00, and announced 1.27.3/1.26.10 at 19:13 PDT — under 7.5 hours end-to-end, with the release process itself taking “just under four and a quarter hours.” The structural reason Go had to move at incident-response speed: on Linux, Go ships its own certificate verifier (crypto/x509) with no revocation checking — a deliberate trade-off that only works when the toolchain can react at browser-vendor speed.

# The standing verification trio for every Go artifact, after any toolchain advisory:

# 1) What toolchain is baked into this binary? (audit trail)
go version -m ./bin/svc-payments | grep 'go1\.'

# 2) Binary-level vulnerability scan (no source required — works on
#    container-extracted artifacts and long-term archives)
govulncheck -mode=binary ./bin/svc-payments

# 3) CI gate: fail the release pipeline if toolchain < current security floor
go version -m ./bin/svc-payments | grep -q 'go1\.27\.3\|go1\.26\.10' \
  || { echo "toolchain below security floor"; exit 1; }

Radar Take: (Opinion.) The durable artifact of this incident is not the patch — it is the URGENT track itself, which converts “rebuild when convenient” into a dated obligation. Encode the trio above into release CI now; trust-store incidents bypass your dependency scanner entirely because the bug is not in your dependency graph. The Go Microservices Production Playbook build conventions are the right place to pin this. If your service layer rides on workflow engines, the same rebuild discipline applies to sidecars and workers — see the Temporal workflow Go architecture series.


6. Quick Hits: Cloud, Models, and the Road to re:Invent

6.1 Cloudflare Developer Platform (Nov 10-13 Changelog)

From the Cloudflare changelog: AI Search can now index D1 database rows as documents (beta, Nov 12) — collapse a vector-ingestion sidecar for small-to-medium stores; Hyperdrive raised connection and configuration limits (Nov 11); Workers/Durable Objects spending limits (Nov 13); Workers gained new Rules hash functions (sha1(), md5(), sha512(), Nov 12); EmbeddingGemma 2 and Inkling 2 Small landed on Workers AI (Nov 12). (Opinion.) AI Search + D1 is the quiet consolidation of “search over your database” into the platform — worth a spike if your RAG ingestion pipeline is a sidecar you pay for.

6.2 GPT-6 Cyber: A “Critical”-Rated Model in Vetted Release

Per OpenAI’s announcement and Reuters coverage (November 5): OpenAI released GPT-6 Cyber, a cybersecurity-specialized model built to find and exploit security flaws, to a small group of vetted customers, rating it “Critical” under its own Preparedness Framework — with reported benchmark figures of 48.5% and reliability rising to 99.93% from 94.49%. (Opinion.) Whatever one thinks of the framing, the operational fact is a new access-control tier (“Daybreak Red”, vetted customers) between frontier capability and the public API — security teams should track who in their org qualifies and what the use policy is.

6.3 Claude Fable 5.2 and Mythos 5.2

Per Anthropic’s announcement (November 9-11): Claude Fable 5.2 (coding, reasoning, agent workloads) and Claude Mythos 5.2 (“most capable models for daunting technical work”) shipped as the week’s flagship pair, targeting coding, finance and the sciences. (Opinion.) The relevant engineering question is not the leaderboard delta but the migration cost: re-run your agent eval suites against 5.2 before pinning, since tool-call formatting and long-context behavior shifts are the usual regression class.

6.4 AWS re:Invent Preview: Durable Functions, Managed Instances, MicroVMs

Per AWS’s official re:Invent news hub (conference Nov 30 – Dec 4, Las Vegas): the serverless track previews new Lambda execution models — Durable Functions, Lambda Managed Instances, and Lambda MicroVMs — with agentic AI infrastructure as the central theme. (Opinion.) “Durable Functions” is the name to watch: durable execution for Lambda would close the gap between step-function orchestration and in-process checkpointing that workflow engines like Temporal (our architecture series) currently own — expect the convergence question at re:Invent.


7. Radar Takeaways: The Monday-Morning List

(Opinion — our synthesis, not the sources’.)

  1. Patch, then evaluate: apply the November 12 PostgreSQL point releases with the full post-update sweep this week; book PostgreSQL 19 as a staging evaluation, not a production upgrade (gate on 19.1).
  2. Retire PostgreSQL 14 now. EOL as of November 12 — every day on 14 is unpatched exposure.
  3. Pin inference floors with verification, not just versions: SGLang ≥ 0.5.22 and vLLM ≥ 0.32.0, with --disaggregation-enable-kv-checksum, KV canaries, the pool-vs-context-length deployment gate, and the two-session canary isolation test in CI.
  4. Formalize the Go rebuild trio (go version -m, govulncheck -mode=binary, CI toolchain floor) — the URGENT track made same-day rebuilds a dated obligation, not an aspiration.
  5. Keep the dual-stack agent wire pinned: A2A v1.0.0 + MCP 2026-07-28 under AAIF governance; watch MCP Dev Summit Nairobi (Nov 19-20) for stateless-core adoption signals.

8. Frequently Asked Questions

PostgreSQL 19 GA on November 19 — should we upgrade production that day?

No. The GA date is scheduled per the official RC2 announcement, and the release notes’ migration section lists steps pg_upgrade cannot automate (locale-provider reindexes for FTS/pg_trgm/fuzzystrmatch, the max_locks_per_transaction default doubling, JIT now off by default). Patch your current major first (the November 12 drop fixed 52 CVEs), evaluate 19 in staging with production-shaped data, and gate production on 19.1. Full playbook: PostgreSQL 19 GA — The Production Upgrade Guide.

What exactly do the KV-cache incidents require us to change this week?

Three concrete items from the published post-mortems: (1) verify engine versions against the floors — SGLang ≥ 0.5.22 (carries the Cloudflare-class fixes) and vLLM ≥ 0.32.0 (GHSA fixes plus hardened defaults); (2) enable transport integrity where supported (--disaggregation-enable-kv-checksum, KV canary detection); (3) add a deployment gate that measures KV pool size against the running engine’s context length — Cloudflare’s pool was sized for 681,280 tokens against a 1,048,576-token context, and the 155 leaked responses were all prompts longer than the pool. Reference: Cloudflare’s remediation post and the VulnCheck advisory for CVE-2026-94627.

Why does the Go eMudhra incident still matter weeks later?

Because the fix only protects rebuilt binaries: Go’s Linux verifier (crypto/x509) performs no revocation checks, so every Go binary that uses TLS and was built before 1.27.3/1.26.10 trusts chains to the compromised intermediates until rebuilt — and Cloudflare confirmed active exploitation October 30. The standing controls are go version -m inventory, govulncheck -mode=binary scanning, and a CI gate that fails releases below the security floor. Reference: the Go blog incident write-up.

Did any agent protocol spec change in the Nov 13-15 window?

No new revision landed — A2A remains v1.0.0 (stable since March 2026, 150+ organizations) and MCP remains at Revision 2026-07-28 (stateless core). What did move: AAIF reached 10 hosted projects on October 22 (Agent Skills, Open Responses, OpenEnv, OpenSandbox joining the six originals), and the MCP Dev Summit in Nairobi (November 19-20) is the next calendar anchor. Treat the stability itself as the signal: it is now safe to pin both specs in integration contracts. Reference: the AAIF press release.

9. Primary Sources

Internal References