Reco Raises $55M and Tencent's Marvis Repositions as AI Agent Sprawl Grows

Reco's funding builds on a $30 million Series B announced in February. AT&T, a customer, invested through its venture arm, along with Forestay and Quadrille Capital. The new round brings total capital raised to $140 million. CEO Ofer Klein said the company's valuation has more than doubled since February and is in the “high hundreds of millions,” while annual recurring revenue is in the “double-digit millions of dollars” and is expected to triple this year. Reco has more than 100 customers, about 40% in financial services. It will use the funds for hiring, sales, partnerships and customer support.

TechCrunch reported that Reco has repositioned from software that maps and secures SaaS and AI platforms to a broader solution that uses a context graph to connect agents to apps, people, accounts and permissions. The platform integrates with more than 280 apps, and Klein said new integrations can be added within days. It uses browser and network signals to find agents outside directly connected apps and has controls to inspect prompts and tool calls. Klein said a Fortune 100 customer found 21,000 agents it did not know about, and a large financial services customer had an agent created by an ex-employee that could access Salesforce and share data with a domain the company could not see.

The market for such tools is crowded. A quick look at public Crunchbase and PitchBook profiles turns up at least two dozen companies selling some form of AI agent security, according to TechCrunch. Some vet tools agents use, some control data access, some build detection and response controls on devices agents run on, and others find unapproved AI usage. Their promises often sound similar, involving knowledge graphs, continuous monitoring, runtime security, tool access and MCP vetting. Security startup HiddenLayer's co-founder and CEO Chris Sestito said earlier this month that when agents reach production, costs and risk scale “really quickly,” and more than 50 of his customers have AI agents in production touching critical systems and sensitive assets. Another startup, Cymphony, said at one U.S. public company it found about 85,000 files accessible to AI tools and agents.

Tencent's Marvis, meanwhile, is taking a consumer and device-level approach. According to Leiphone, IDC estimates there will be 79.4 million active agents globally in 2026, up nearly three times from the previous year. As agents spread across PCs, they leave behind generated files, browser data, downloads, installed software, accounts and local data. Marvis positions itself as an AI butler that manages these digital assets rather than completing single tasks. Its ecosystem head introduced the concept of “AI service direct” at Arm's conference in early September. Tencent vice president Lin Songtao said at WAIC that AI applications are moving from answering questions to delivering results, and that without result delivery there is no business loop.

Marvis divides roles: agents are helpers that understand, execute and deliver on a per-task basis, while an AI butler is a manager that long-term understands a user's computer, files, devices and digital assets. In its sidebar, Marvis offers a “butler team”: a computer butler for performance and troubleshooting, a file butler for knowledge bases, a software butler for applications and a browser butler for web automation. Game and life butlers are planned. The computer butler was built with Tencent's enterprise IT team, internally called “8000,” which has served tens of thousands of Tencent employees over more than a decade. Their Skill list covers four categories—operating systems, hardware, software and networks—and 27 real operations scenarios. It translates colloquial complaints into diagnostic paths, using system logs and historical cases. Hardware and peripherals are the most heavily covered area.

The file butler keeps files in place and automatically clusters them, rather than requiring users to upload and categorize. It connects to Tencent Docs and plans to add WeCom, and a “project” feature will gather tasks and outputs in one context. The software butler moves from managing applications to managing services. On gaming, Marvis has worked with Microsoft, Intel and Qualcomm on system- and hardware-level tuning, and a Steam AI teaching demo provides battle analysis and real-time key prompts. A game butler will launch later. On life services, a Luckin co-brand drew more than 200,000 participants and over 10 million total exposures, and more than 50 service providers are integrating. Marvis requires password confirmation for system-level changes and does not bypass security mechanisms.

Marvis is also working across devices. A phone can initiate a request while a PC or NAS executes locally, with cloud compute added when needed. Marvis indexes local files, applications and system state on-device so agents have context without starting from a cloud search. For PC and NAS makers, building full agent capabilities requires long-term model, engineering and product investment; connecting to an existing agent engine can add AI capabilities on top of current hardware and systems. For Marvis, entering OEM and Xinchuang devices means becoming part of the device rather than only a consumer app. It already works with Microsoft and Intel and is pursuing OEM and Xinchuang partnerships.

Cost and privacy are handled partly through on-device models. Through low-bit quantization, Marvis has compressed a 35B-level model from about 30GB to under 10GB so it can run locally, keeping data off the cloud and reducing cloud costs. For deletion, modification and payment operations, Marvis explains and then asks for confirmation. External services follow a “default disagree” principle with dual developer and user authorization. Tencent's Application Bao business previously distributed software and apps; in the agent era, Marvis is trying to complete tasks for users across PCs, NAS, AI boxes and eventually cars and glasses.