Developers stopped writing code by hand. Amazon told 80% of its engineers to use AI tools weekly. Google released a browser-native version of MCP. And Claude Opus 4.6 set a new agentic benchmark at 65.4% on Terminal-Bench 2.0. Here is what mattered most in AI this week.
AI Coding Tools: Developers Who Don't Write Code
Amazon set a firm target this week: 80% of its developers must use AI coding tools at least once per week. The directive puts a number on what other tech giants are already seeing in practice. Spotify's best engineers have not written a single line of code since December. Claude Code handles remote deployment. Engineers push changes to production from their phones during their commute.
Anthropic's Boris Cherny, who leads Claude Code, said he personally has not written code in over two months. Anthropic itself now generates 70% to 90% of its code through AI. OpenAI made a similar claim, saying GPT-5.3-Codex "was instrumental in creating itself." These are not distant predictions. They are current operational realities at the companies building the tools.
The February 2026 model rankings tell a parallel story. Claude Opus 4.6 launched on February 5 and scored 65.4% on Terminal-Bench 2.0, a benchmark designed for multi-step agentic engineering tasks. Developers evaluating tools no longer ask which model is smartest. They ask which one gets the job done on the first pass and fits their workflow without constant correction.
The productivity evidence remains mixed in the broader market. GitHub Copilot crossed 20 million users in mid-2025, a 400% jump in one year. But a METR study found that experienced developers who believed AI made them 20% faster were objectively 19% slower in controlled tests. Amazon's push suggests enterprise leadership believes the long-term gains are real, regardless of what individual benchmarks show.
AI Processing: MCP Security Cracks Start to Show
The Model Context Protocol has become the connective tissue for agentic AI. But this week, three new CVEs landed against Anthropic's Git MCP server: CVE-2025-68145, CVE-2025-68143, and CVE-2025-68144. The flaws enable remote code execution through prompt injection, a path validation bypass, unrestricted git_init, and argument injection. Microsoft's MarkItDown MCP server showed similar vulnerabilities in the same report.
SD Times ran a detailed breakdown on February 16 of MCP's structural privacy and security problems. Data leakage is the core risk. An AI agent with role-based access controls can still infer data it does not have permission to read. A salesperson's AI copilot, for example, can accurately predict profit margins on products it cannot directly access, then inject that information into a customer document.
The MCP ecosystem now has over 30 official server implementations. Past incidents this year include a malicious MCP server that exported WhatsApp history in April, a prompt-injection attack against GitHub's MCP server in May that pulled data from private repos, and an Asana bug in June that let organizations see other organizations' data.
The response from the security community is to treat MCP servers as untrusted by default and apply rigid sandboxing around every tool execution. Companies like Tray.ai are building control planes that sit between MCP servers and agent execution layers. Anthropic contributed MCP to the Linux Foundation's Agentic AI Foundation in early 2026. Openly governed standards increase scrutiny, and right now, that scrutiny is finding real problems.
Despite the vulnerabilities, adoption continues to accelerate. Google released an early preview of WebMCP on February 13, a browser-native extension of the protocol. WebMCP gives AI agents structured tools to perform actions on websites, from purchases to complex bookings, without needing to parse the DOM. The announcement signals that Google is betting on MCP as the operating layer for the agentic web.
Standards and Protocols: WebMCP and the Agentic Web
Google's WebMCP launch this week is the clearest signal yet that the open web is reorganizing around AI agent interaction. Traditional web bots guess their way through page structure by analyzing the DOM, a method that breaks easily when code changes. WebMCP gives AI agents a declared, structured interface to act within websites rather than just read them.
The Agentic AI Foundation, formed under the Linux Foundation after Anthropic's MCP contribution, is now working on multi-agent orchestration standards. The next phase of MCP goes beyond a single model connecting to many tools. The AAIF is designing standards for multiple agents, each with their own MCP servers, to collaborate on complex projects. A marketing agent might use a creative suite MCP to generate an ad, then hand it to a legal agent connected to a compliance database for review.
IBM's Kate Blair said this month that 2026 is the year multi-agent systems leave the lab and enter production. IBM contributed its BeeAI and Agent Stack projects to the Linux Foundation, and Blair noted that openly governed community standards are what will unlock broader ecosystem participation. The A2A protocol from Google is approaching its first major release. MCP, A2A, and IBM's ACP are converging around shared governance rather than competing in isolation.
For data teams, this convergence matters directly. Dremio's MCP server is already available, and as agent orchestration standards mature, the ability to connect AI agents to a governed, query-optimized lakehouse becomes a first-class engineering concern rather than a custom integration project.
Experience the Future with Dremio
The AI landscape changes fast. Data teams need tools that keep pace.
Dremio's semantic layer and Apache Iceberg foundation let you build AI-ready data products. The platform handles optimization automatically. You focus on insights, not infrastructure.
Ready to see agentic analytics in action? Start your free trial today and experience the autonomous lakehouse.
Top comments (0)