Google expands Gemini Flash and managed agents with computer-use and background tasks

What changed: Google introduced Gemini 3.6 Flash and expanded the Flash family, while Gemini 3.5 Flash gained the ability to control a computer and use apps directly for multi-step work. Gemini API managed agents now support background task execution and remote Model Context Protocol (MCP) connections, and Google outlined an AI Control Roadmap for how production agents should be monitored and constrained.

Why it matters: These upgrades move Gemini from a chat interface toward full workflow automation, letting agents click through interfaces, keep long-running jobs alive, and plug into external tools through MCP. Founders and operators get a clearer pattern for building agents that do real work on user machines while staying inside governance guardrails.

Try/watch: If you are already using Gemini, experiment with a narrow, high-value agent that runs in the background (for example, weekly reporting or ticket triage), and design explicit guardrails aligned with Google's control recommendations to avoid silent failures or runaway actions.

MCP spec goes stateless, cutting agent infrastructure overhead

What changed: The Model Context Protocol (MCP), a standard for connecting AI agents to tools and data sources, shipped a new spec dated 2026-07-28 that shifts the protocol to a stateless architecture, removing session identifiers and the traditional initialize handshake. The update also adds authenticated communication between agents and MCP servers, self-describing requests via a meta object, header-based routing for streaming HTTP, and multi-round-trip requests that replace custom server-to-client query flows. GitHub reported fully removing Redis databases previously used for MCP session management, sharply reducing read/write operations in its infrastructure.

Why it matters: Stateless MCP makes it cheaper and simpler to scale agent backends, especially when you are orchestrating many short-lived tools or micro-agents. Stronger authentication and structured metadata help enterprise teams meet security and observability requirements without building a custom protocol.

Try/watch: If your agents already use MCP or a similar pattern, start planning a migration to stateless operations and per-request authentication, and instrument meta fields so you can trace which agent, tool, and tenant triggered each call across your stack.

Meituan open-sources LongCat-2.0 and details self-evolving fulfillment agents

What changed: Meituan released LongCat-2.0 as open source, a 1.6-trillion-parameter mixture-of-experts model with about 48 billion active parameters per forward pass, designed specifically for complex agentic coding tasks. The company’s fulfillment AI team presented research at ACL 2026 on a self-evolving agent operating system that combines continuous pre-training, post-training, agentic reinforcement learning, and multimodal understanding to power logistics agents. The same report highlights analysis of 3,607 user-reported AI agent incidents from 2025–mid-2026, showing overeagerness and misalignment as the top failure modes, each appearing in more than 43% of cases. Anthropic also launched Claude Opus 5 on July 24, a frontier model positioned as the new default for Claude Max and Pro that matches its highest-end model on coding tasks at about half the cost.

Why it matters: LongCat-2.0 gives builders a powerful, open-weight option for agentic coding, while Meituan’s fulfillment stack shows how to turn LLMs into a continuously improving operations platform rather than one-off bots. The incident data and Opus 5’s cost profile underscore a practical trend: production agents must be tightly scoped, monitored for behavior drift, and powered by models that balance frontier performance with predictable economics.

Try/watch: Consider prototyping agentic coding workflows on LongCat-2.0 or Claude Opus 5, but pair them with explicit incident logging categories (overeagerness, misalignment, latency, etc.) so you can quantify failure patterns instead of relying on anecdotal feedback.

Agentic commerce moves from pilots to live B2B and on-chain payments

What changed: Visa and Lianlian completed Greater China’s first live business-to-business transaction executed via agentic systems, showing agents can handle real payment flows between firms. Samsung began shipping Gemini Intelligence on new Galaxy devices, allowing on-device AI to automate tasks all the way up to the payment step for consumer workflows. In parallel, Coinbase launched Agentic Wallets that let AI agents hold funds and send on-chain payments without human approval, and Visa and Mastercard confirmed their first end-to-end financial transactions executed entirely by AI agents.

Why it matters: These moves signal that payments is becoming a first-class agent domain rather than a demo, with both consumer and B2B flows now testing autonomous execution. Builders of SaaS tools, marketplaces, and fintech products can start treating agents as transaction participants, not just recommendation engines, but will need to invest heavily in policy, auditing, and kill switches.

Try/watch: If you operate in commerce or fintech, map a single low-risk payment workflow (for example, small recurring supplier payouts or refund handling) where an agent could prepare and propose the transaction, and design a staged rollout that keeps a human in the loop while you gather data on accuracy, fraud patterns, and user trust.