AI News Roundup: October 19-25, 2026 — TanStack AI 1.0 Goes Stable, Four Projects Join the Agentic AI Foundation, and Agent Identity Hits the Mainstream
The week of October 19-25 delivered three structurally significant stories at once. TanStack AI 1.0 shipped as the first stable, type-safe, provider-agnostic TypeScript agent stack — closing the "JavaScript agent framework" gap that has existed since the OpenAI Agents SDK went v2. The Agentic AI Foundation accepted four donated projects from Anthropic, OpenAI, Hugging Face, and Alibaba, moving the core agent protocols under neutral governance for the first time. And a US national-security access order against Anthropic's most advanced models — paired with proposed expansion of AI and semiconductor export restrictions — turned the three-way local-runtime pattern from an interesting option into an enterprise hedging requirement. Meanwhile, ServiceNow, Microsoft, and WorkOS all shipped or explained agent-identity features within 72 hours of each other, making "cryptographic agent identity" this week's fastest-moving mainstream theme.
TanStack AI 1.0: The TypeScript Agent Stack Grows Up
TanStack AI 1.0 (released Friday, October 23) is the biggest agent-framework release of the month. The v1.0.0 release marks every core API as stable, removes all previously deprecated 0.x APIs, and lands as a monorepo of 76 packages covering the full agent-application surface: streaming chat, tool-calling agents, structured outputs, realtime voice, media generation, and MCP server support.
The headline features:
- Typed agents with a built-in agent loop — agents, subagents, and coding-agent workflows are first-class, with per-model typing for provider-specific capabilities
- 23 provider adapters — OpenAI, Anthropic, Gemini, and twenty more, behind one type-safe interface; the provider-agnostic promise extends to switching mid-project without rewriting tool definitions
- Portable agent skills — skills move between agents and environments, aligning with the Skill Protocol line the ecosystem has consolidated around this year
- Computer-use tooling and sandboxes — the coding-agent workflows ship with sandboxed execution out of the box
- Experimental multi-environment harness — a toolkit for running agents across multiple environments (local, remote, cloud) with a single harness API; marked experimental in 1.0
- Durability and interrupts — long-running agent state survives process restarts, and interrupts are a native primitive rather than a bolt-on
- Isomorphic tools — the same tool definition works client- and server-side with shared type safety
- Framework hooks for React, Vue, Svelte, and Solid — chat UI primitives that don't lock you into one frontend stack
The release blog post, titled "Still Your AI, Still Your Way," positions TanStack AI explicitly against provider lock-in — a message that landed the same week an access order demonstrated exactly what lock-in risk looks like in practice.
An open RFC on composable AI plugins (GitHub Discussion #954) is exploring a plugin architecture for apps built on chat(), generateImage(), and generateSpeech() primitives — a signal that TanStack intends to be the substrate, not just another SDK.
Where It Sits
| TanStack AI 1.0 | Vercel AI SDK | OpenAI Agents SDK (TS) | |
|---|---|---|---|
| Provider adapters | 23 | Broad | OpenAI-first |
| Agent loop | Typed, built-in | Via agents package | Core primitive |
| MCP support | Server + client | Yes | Yes |
| Durability/interrupts | Native | Partial | Via extensions |
| Frontend hooks | React/Vue/Solid/Svelte | React-first | None |
| Sandboxes | Built-in | Via partners | Via partners |
The strategic read: the Vercel AI SDK owns the Next.js-native slot, the OpenAI Agents SDK owns the OpenAI-native slot, and TanStack AI 1.0 is making a credible play for the framework-agnostic, type-safety-first slot — the "TanStack Query of agents" position. Given that TanStack Query became the default data-fetching library in React precisely by being framework-agnostic and type-safe, the playbook is proven.
The Agentic AI Foundation: Four Donations, Neutral Governance
On October 24, the Agentic AI Foundation announced it had taken over stewardship of four major projects, bringing its hosted portfolio to ten:
- Anthropic's Agent Skills — the skills format behind the Skill Protocol line
- OpenAI's Open Responses — the open reference implementation of the Responses API wire format
- Hugging Face's OpenEnv — the open agent-environment interface
- Alibaba's OpenSandbox — the sandboxed-execution standard used by coding agents
This is the governance story of the quarter. Throughout 2026, the agent ecosystem's core primitives — the Responses wire format, the skills format, the environment interface, sandboxing — were owned by the same companies that sell hosted models on top of them. Moving them to a neutral foundation changes three things:
- Protocol evolution decouples from vendor roadmaps. MCP 2.2's cross-server chaining shipped fast because Anthropic pushed it; future protocol work now has a venue where OpenAI, Anthropic, Alibaba, and Hugging Face are contributors rather than owners.
- The three-way local-runtime pattern gets an anchor. Mistral Agent Runtime 2.0's OpenAI wire-format parity was a unilateral compatibility move; with Open Responses under neutral governance, wire-format parity becomes a maintained standard rather than a vendor's goodwill.
- Procurement gets simpler. Enterprises evaluating agent stacks can now point to foundation-governed specs for the four riskiest layers (protocol, skills, environments, sandboxing) instead of four different corporate CLAs.
The timing is not accidental: it lands in the same week as the access-order shock (below), and the implicit message is that the substrate survives even when any single vendor's access does not.
The Access-Order Week: Anthropic Shutdown, Export Expansion, and the Local Hedge
The most consequential news of the week is regulatory, not technical.
Anthropic cut off access to its most advanced models following a Trump-administration order citing national-security concerns (Bloomberg). The exact scope remains fluid as of this writing, but enterprise customers of the affected tiers began receiving access-transition notices, and the incident immediately rewrote procurement checklists: "what happens to our agents if our model vendor's access is ordered off" went from a theoretical row in the risk register to a live question with a deadline.
In the same window, proposed regulation expanding AI and semiconductor restrictions advanced (reported by The Information alongside its Kimi K3.1 Pro coverage), extending the export-control logic that China's AI-agent export controls established earlier in the year — now mirrored on the US side.
Two model-roadmap signals completed the picture:
- GPT-6.1 Astra was cancelled. The WSJ reported that OpenAI scrapped the anticipated release after internal safety tests; OpenAI did not ship GPT-6.1 Astra this month. Enterprise buyers who had pinned Q4 plans to Astra are re-evaluating, and the cancellation strengthens the position of GPT-5.6 Sol Enterprise as the stable OpenAI enterprise tier for the remainder of 2026.
- Moonshot AI launched Kimi K3.1 Pro (October 23) — a frontier-tier launch that landed, notably, in the same news cycle as the restriction expansion. Kimi K3.1 Pro is now the most capable model in the Kimi line, and Moonshot's enterprise pitch (Kimi Code CLI plus hosted K3.1 Pro) is squarely aimed at teams that want a non-US, non-sanctioned production path.
The Local Hedge Becomes Policy
The three-way local-runtime pattern documented in the October 23 roundup — Claude Code, OpenAI Codex, Mistral Agent Runtime 2.0, with Qwen3.9 Flash as the local-model entry — was framed then as an interesting portability property. This week it became a procurement requirement. The pattern enterprises are writing into playbooks:
| Layer | Hosted bet | Local hedge |
|---|---|---|
| Orchestration | Claude Code / Codex | Mistral Agent Runtime 2.0 |
| Model | Claude Opus 4.7 / GPT-5.6 Sol | Qwen3.9 Flash (Apache 2.0), Kimi K3.1 Pro (API) |
| Tools | MCP 2.2 (hosted gateways) | Local MCP gateway + OpenSandbox |
| Wire format | Responses API | Open Responses (now foundation-governed) |
The Agentic AI Foundation's timing now reads as coordinated rather than coincidental: every layer of the local hedge column is now either open source or foundation-governed.
Agent Identity Goes Mainstream
The identity thread this site has tracked all year — SwarmAuth GA, Rip-PR's signed reviews, Cogni's writer identity — crossed into the mainstream this week, with three announcements inside 72 hours:
ServiceNow AI Passport (Oct 19)
AI Passport is a "cryptographically signed runtime credential" for agent identity and scope enforcement. Each agent gets a signed passport encoding its identity, permitted scopes, and validity window; services verify the passport before honoring agent requests. This is the SwarmAuth pattern shipped by a major enterprise platform vendor rather than an open-source project — the clearest sign yet that signed agent identity is becoming table stakes.
Microsoft Copilot Studio Gets Entra Identities (Oct 19)
Preview Copilot Studio agents are gaining Entra identities, mailboxes, and Teams presence. An agent with an Entra identity is an agent that can be audited, scoped, and deprovisioned through the same machinery as employees — which is precisely what IT departments have been asking for since agents started touching email and calendars. Giving agents mailboxes and presence also normalizes agents as first-class organizational actors, with everything that implies for accountability.
WorkOS on Agent-to-Agent SSO (Oct 21)
WorkOS published a widely-circulated explainer on agent-to-agent SSO and federation gaps, riding Okta's Agent SSO momentum. The core problem: human SSO assumes a human presenting a credential; agent-to-agent calls have no session, no browser, and often no stable identity. The piece walks the federation gaps and emerging patterns (service-identity tokens, delegated agent credentials) — and its readership numbers suggest the market is now actively shopping for solutions rather than theorizing.
Why This Week Mattered
Signed agent identity has moved through the standard adoption curve faster than any agent-infrastructure theme this year: research concept (spring) → open-source implementations (SwarmAuth, Rip-PR, summer) → GA products (fall) → enterprise platform features (this week). The Wikipedia article on the July 2026 rogue-agent incidents remains the reference case everyone cites for why this layer exists.
Long-Running Agents: Notion's 24-Hour Harness and Daytona's $1B Round
Two announcements extended the "agents that run for days" thread:
Notion's Sixth-Generation Agent Harness (Oct 20)
Notion shipped a sixth-gen agent harness in which tasks can exceed 24 hours — up from five minutes in the previous generation. The upgrade bundle: an Auto model router that moves work across model tiers as task phases change, two-tier credits separating agent execution from tool usage, and enterprise BYOK (bring your own key). The five-minutes-to-24-hours jump is the story: it reflects both durability improvements (checkpointing, resumable state) and a user base that now trusts agents with genuinely long work — annual-report assembly, multi-stage data cleanup, cross-workspace migrations.
Daytona Raises $60M Series B at $1B Valuation (Oct 24)
Daytona raised a $60M Series B at a $1B valuation for sandboxed, long-running agent infrastructure. Daytona's pitch is the compute substrate for exactly the workloads Notion's harness implies: agents that need secure, persistent, isolated environments for hours or days, not seconds. The round is also the clearest venture signal yet that "agent infrastructure" is being capitalized as its own category — distinct from model providers, frameworks, and observability.
Also in this thread: Google Cloud expanded VPC Service Controls perimeter controls for agents (Oct 24), closing the exfiltration paths that long-running agents with standing credentials otherwise open up.
Government Agents: Gemini for Government Services
Google Cloud previewed Gemini for government services (Oct 21) with built-in skills and partner agents for FOIA processing and redaction. The significance is less the product than the workload: FOIA and redaction are document-heavy, deadline-bound, audit-required processes — the exact shape of work where 24-hour agent runs with signed identities and per-scope credentials (see above) are the enabling stack. Every layer this site has tracked this year — long-running harnesses, agent identity, guardrails, MCP-governed tool access — shows up in the government-agent reference architecture.
Security: F5 Agent Gateway and Cisco's Agentic FAITH
- F5 + Google Cloud detailed an Agent Gateway + AI Guardrails integration (Oct 22) for securing agent fleets at the network layer — traffic policing and guardrail enforcement in front of agent workloads rather than inside agent code.
- Cisco launched Agentic FAITH with the VulnLoc benchmark (Oct 24), a framework for evaluating multi-step security agents — measuring whether security agents can locate, verify, and remediate vulnerabilities across multi-step investigations without drifting. A benchmark specifically for multi-step security agents is new; prior security-agent evaluation stopped at single-step tool use.
Both fit the site's formal-safety thesis: the AgentGuard (OWASP-native) and OpenShell (property-based) approaches now have network-layer and benchmark-layer counterparts. The stack is filling in.
Economics: Gartner's Fivefold Cost Forecast
Gartner predicted a fivefold increase in AI agent costs despite cheaper tokens (circulated October 23). The argument: token prices keep falling, but agent workflows consume orders of magnitude more tokens than chat did — multi-step loops, long-horizon tasks, 24-hour harnesses, retrieval over entire corpora — so total spend per agent workflow rises even as unit cost falls. The forecast lands in the same week as Notion's two-tier credits (metering agent execution separately from tool usage) and Daytona's round (capitalizing the execution layer). The cost story of 2027 is being written now, and it is a metering and routing story, not a token-price story.
Also Notable
- Meta launched its Muse agent in Japan — Muse's first expansion beyond North America, with Gmail integration among the launch features. Consumer agent expansion into Asia is now officially underway.
- Kognitos launched 61 role-based AI Employees for finance, backed by 515 encoded SOPs — the most aggressive "agents as headcount" SKU to date.
- OpenAI, Anthropic, and Google DeepMind's safety collaboration (Reuters, ongoing since September) continues as background context for the access-order week; the three-lab cooperation line is one of the few steady signals in an otherwise volatile regulatory picture.
Website Content Added This Period
New Agent Entries
- TanStack AI — the type-safe, provider-agnostic TypeScript agent SDK, 1.0-stable with 23 provider adapters and MCP support
- Kimi K3.1 Pro — Moonshot AI's new frontier-tier model, launched October 23
- Daytona — sandboxed, long-running agent infrastructure, $60M Series B at $1B valuation
Themes for the Week
- The TypeScript agent stack is complete. TanStack AI 1.0 gives JavaScript/TypeScript a stable, type-safe, provider-agnostic agent substrate — the last major ecosystem gap in agent frameworks.
- Neutral governance arrived for the core protocols. Agent Skills, Open Responses, OpenEnv, and OpenSandbox under the Agentic AI Foundation means the substrate no longer depends on any single vendor's roadmap or access policy.
- Access risk is now a procurement input. The Anthropic access order and the GPT-6.1 Astra cancellation make "local hedge" a line item, not an ideology — and every layer of the local stack is now open or foundation-governed.
- Agent identity went mainstream in 72 hours. ServiceNow AI Passport, Copilot Studio Entra identities, and WorkOS's federation explainer move signed agent identity from open-source frontier to enterprise platform feature.
- Agents run for days now. Notion's 24-hour harness and Daytona's $1B round define the long-running-agent layer; durability is no longer exotic.
- Cost is the 2027 story. Gartner's fivefold forecast, Notion's two-tier credits, and model-router features across the board: agent economics is now a metering and routing problem.
Looking Ahead
- Wire-format consolidation under the Foundation — does OpenAI keep Open Responses as a first-class supported target now that it's foundation-governed? The answer shapes the three-way runtime pattern's durability.
- Kimi K3.1 Pro enterprise traction — Moonshot's timing (launching into the access-order week) gives it a real opening in non-US-pinned enterprise stacks; watch for framework adapters.
- TanStack AI harness going stable — the experimental multi-environment harness is the piece that would make TanStack the substrate for cross-environment agent testing; the AI Plugins RFC suggests the ecosystem agrees.
- Identity interop — SwarmAuth, AI Passport, and Entra identities solve the same problem three ways; the first bridge between them is the next unlock.
The site will continue tracking these threads through Q4 2026, with framework-level entries added as the new tools stabilize.