AI Agent Digest: Week 25, 2026 - Salesforce Buys Fin for $3.6B, Microsoft's Office Agents Start Doing, China Open-Sources a Frontier Model

This was the week the agent market stopped pitching and started buying. Salesforce wrote a $3.6 billion check for a customer-service agent, Microsoft quietly flipped its Office agents from "suggest" to "do," and a Chinese lab open-sourced a frontier model just to make the whole stack cheaper. Here are the eight stories that actually moved the needle, with our take on what each one means.

1. Salesforce buys Fin (formerly Intercom) for $3.6 billion

Salesforce signed a definitive agreement to acquire Fin, the AI customer-service platform spun out of Intercom, for $3.6 billion (TechCrunch, CNBC). Fin's agent resolves an average of 76% of support volume end to end across chat, email, WhatsApp, SMS, phone and Slack, and brings a 30,000-company customer base into the Agentforce fold. It is the largest agentic CX deal to date.

Hot take: This is the moment "AI agent" stopped being a feature and became an asset class. Salesforce did not buy a model or a research team. It bought resolution rates and a customer base, because that is what an agent business actually is now. Every SaaS incumbent watching this is doing the same math: build the agent layer or buy it before a competitor does. Expect more nine-figure agent acquisitions before Q4.

2. Microsoft's Copilot agents go from suggesting to doing

Microsoft pushed its agentic Copilot features into general availability across Word, Excel and PowerPoint. The assistant now actively edits documents, restructures spreadsheets and builds slide decks rather than politely proposing changes in a sidebar (artificialintelligence-news).

Hot take: The gap between "Copilot that suggests" and "Copilot that does" is the entire ballgame, and Microsoft just crossed it for hundreds of millions of seats. Suggestion engines are toys. Agents that take the action are coworkers. The catch nobody is saying out loud: once the agent edits your spreadsheet directly, the trust bar goes from "is this a good idea" to "did it just break my model," and that is a far less forgiving standard.

3. NVIDIA and ServiceNow ship Project Arc, a self-evolving desktop agent

At ServiceNow Knowledge 2026, NVIDIA and ServiceNow expanded their partnership to deliver governed autonomous agents for enterprises, headlined by Project Arc: a long-running, self-evolving desktop agent built for knowledge workers (Shakudo).

Hot take: "Self-evolving" and "governed" in the same sentence is the tension the entire enterprise category is wrestling with. A desktop agent that learns and persists is exactly what makes it useful, and exactly what makes a CISO nervous. The winners here will not be the most capable agents. They will be the ones that can show you a clean audit trail of everything they changed and why.

4. Zhipu AI open-sources GLM-5.2 under MIT license with a 1M-token context

Chinese lab Zhipu AI announced it will open-source GLM-5.2 under the permissive MIT license, with a one-million-token context window, a direct counter to tightening US chip restrictions (dentro.de/ai).

Hot take: Every time a frontier-class model goes MIT-licensed, the floor under proprietary agent pricing drops a little further. A million-token context that costs nothing to license is a gift to anyone building agents with long memory, which is most serious agent work. The geopolitical subtext is the real story: export controls meant to slow China down are accelerating the open-weights ecosystem that undercuts everyone's margins.

5. Anthropic's Project Glasswing turns agents loose on zero-days

Anthropic gave select organizations including AWS, Apple, Cisco, Google, JPMorgan Chase and Microsoft access to its Mythos preview through Project Glasswing, a program aimed at finding critical software vulnerabilities. In weeks of testing the model surfaced thousands of zero-day vulnerabilities (crescendo.ai).

Hot take: This cuts both ways and pretending otherwise is naive. An agent that finds thousands of zero-days for defenders is the same agent that finds thousands of zero-days for attackers. Anthropic keeping it inside a 50-partner walled garden is the right call for now, but the capability is out of the bottle conceptually. Security teams that are not already planning for agent-discovered exploits are planning to lose.

6. JPMorgan reclassifies AI from R&D to core infrastructure

JPMorgan Chase formally moved its AI investments out of the experimental R&D column and into core infrastructure, backed by a roughly $19.8 billion 2026 technology budget and 2,000 staff dedicated to AI development (crescendo.ai).

Hot take: Accounting changes are boring until you realize what this one signals. "Core infrastructure" means AI agents are now treated like the mainframes and trading systems the bank cannot operate without, not like a science project that might get cut next downturn. When the most conservative institution on Wall Street says agents are load-bearing, the "is this real" debate is over.

7. Agentic payments grow up: MetaMask, Ant International and Fireblocks

Payment rails for agents matured fast this month. MetaMask shipped an Agent Wallet that lets agents execute onchain trades across EVM chains under mandatory security checks, Ant International introduced the Agentic Mobile Protocol for wallets and super apps, and Fireblocks launched an Agentic Payments Suite for agent-initiated stablecoin payments on any blockchain (PYMNTS, Payment Executive).

Hot take: An agent that can research and draft is interesting. An agent that can actually pay is a business. The infrastructure landing now is what turns agents from advisors into operators, and it is also where the scariest unsolved problem lives: when an agent makes a bad purchase or gets exploited, nobody has agreed who eats the loss. Capability is shipping faster than liability law, as usual.

8. The funding bubble nobody wants to name

The money is rotating hard. VCs are pulling out of thin application wrappers and pouring into the infrastructure layer: multi-agent orchestration, security and compliance, and inter-agent communication. Startups without proprietary routing engines or enterprise-grade governance are reportedly facing quiet down-rounds, and a meaningful slice of early agent startups may exhaust their runway by late 2026 under heavy token costs and slow enterprise sales cycles (Product Leaders Day).

Hot take: The "wrap an API and call it an agent" era is ending, and good riddance. What survives the shakeout is what should have mattered all along: real memory, real governance, real ability to execute work end to end. The startups dying are the ones that confused a clever demo with a durable product. The ones living owned the boring infrastructure underneath.

What we're watching next week

• Whether a second SaaS incumbent answers Salesforce with its own nine-figure agent acquisition.
• How fast Microsoft's "agents that act" rollout collides with enterprise change-management and audit requirements.
• Any movement on agentic-payment liability standards, the gap that could stall the whole commerce wave.

Bottom line

Week 25 was about commitment. Acquisitions, GA rollouts, frontier open-weights and accounting reclassifications are not the language of hype. They are the language of organizations that have decided agents are permanent and are now arguing about governance, not existence. The question stopped being whether to deploy agents and became which ones you can actually trust to act on your behalf.

That last word, trust, is the whole game. An agent you cannot audit is a liability, not an employee. That is exactly the bet behind Geta.Team: AI employees with persistent memory, their own identity, and a transparent record of what they do, self-hosted so your data stays yours. If this week proved agents are here to stay, the next question is whose you can actually trust. Come find out.