On Monday, Anthropic's Claude Sonnet 4.6 scored 72.5% on OSWorld — navigating spreadsheets, filling multi-step web forms, and coordinating across browser tabs with near-human accuracy. The same week, Fujitsu shipped an AI platform that compressed three person-months of medical software development into four hours. Meanwhile, Microsoft quietly disclosed that only 15 million customers actually pay for Copilot — 3.3% of its 450 million Microsoft 365 subscribers.
February 2026 is the month AI agents crossed the capability threshold from useful assistants to functional workers. The models can operate computers. They can write, test, and ship code. They can control mobile applications and analyze two-hour videos. But the Copilot number tells the real story: capability and deployment are on divergent trajectories, and the gap is widening.
This is why AgentPMT exists. The problem was never whether agents could do the work — it's whether enterprises have the infrastructure to deploy them with accountability, cost transparency, and cross-platform flexibility. AgentPMT's drag-and-drop workflow builder gives every agent clear tasks and defined boundaries. Dynamic MCP connects agents to the largest marketplace of AI tools and AI skills through a single integration point, without bloating context windows. And complete cost transparency — per-tool pricing, per-workflow cost tracking, hard-enforced budget controls — means enterprises know exactly what every agent interaction costs before the invoice arrives. The agents are ready. The question is whether your infrastructure is ready for the agents.
The Capability Threshold Just Fell
The model race has shifted from "smarter" to "more capable as workers." OSWorld and SDLC automation aren't benchmarks for researchers — they measure whether agents can replace the exact tasks that fill most knowledge workers' days.
Sonnet 4.6, released February 17, delivers a 72.5% OSWorld score — navigating spreadsheets, filling multi-step web forms, coordinating across browser tabs. It doubles the context window to one million tokens while maintaining the same pricing as Sonnet 4.5 at $3 and $15 per million input and output tokens. Anthropic made it the default model for all Claude users, free and paid. MCP connectors now work inside Excel, connecting Claude to S&P Global, LSEG, Daloopa, PitchBook, Moody's, and FactSet without leaving the spreadsheet.
The same day, Fujitsu announced an AI-driven software development platform that uses multi-agent collaboration across the entire software development lifecycle — from requirements analysis to integration testing. The headline number: medical software modifications that previously took three person-months were completed in four hours. Fujitsu plans to expand the platform to finance, manufacturing, retail, and public services.
Alibaba released Qwen 3.5 on February 16 with visual agentic capabilities for mobile and desktop application control. Bloomberg reported the model is 60% cheaper and handles large workloads eight times better than its predecessor, with the ability to analyze videos up to two hours long. OpenAI shipped GPT-5.3-Codex-Spark on February 12, running on Cerebras chips at over 1,000 tokens per second — 15 times faster than its flagship Codex and the first OpenAI model running off Nvidia hardware entirely.
Multiple labs, same week, all shipping agents that operate computers and development environments autonomously. This isn't incremental improvement. This is a category change.
AgentPMT's Dynamic MCP ensures that agents like Sonnet 4.6 can access the tools they need at the moment of execution — without preloading hundreds of tool definitions into context. With one-million-token context windows now standard, the question shifts from "does the agent have enough context?" to "is the agent wasting context on tool definitions instead of actual work?" Dynamic MCP's remote tool fetching means the agent searches for what it needs, pulls in only that tool's schema, executes it, and moves on. The hundreds of other tools in the marketplace never touch the context window. As we demonstrated in our research, MCP servers can waste 96% of agent context on tool definitions — Dynamic MCP eliminates that problem entirely.
The Deployment Gap Is Getting Wider, Not Narrower
The agents can use computers. Enterprises can't use the agents. And the data keeps getting worse.
Microsoft disclosed in its February 9 earnings that Copilot has 15 million paid users — 3.3% of its 450 million Microsoft 365 subscribers. That's 160% year-over-year growth, but the absolute penetration number is brutal for the most widely distributed AI product in the world, priced at $30 per user per month.
Deloitte's State of AI in the Enterprise 2026 report paints a broader picture. Twenty-three percent of enterprises are using agentic AI, and 74% plan to adopt within two years. But only 11% have agents in production. Only 21% have mature governance frameworks. The ambition-to-execution gap is enormous.
Salesforce's eleventh annual Connectivity Benchmark, conducted with Vanson Bourne and Deloitte Digital across 1,050 IT leaders, adds the operational detail. The average enterprise runs 12 agents. Fifty percent operate in silos. Only 27% of applications are integrated. Ninety-six percent of IT leaders say integration is critical to agent success, but 86% fear agents will add more complexity to their already fragmented tech stacks. Gartner still predicts 40% of agentic AI projects will be cancelled by 2027 due to security and governance failures.
The pattern is clear: as agent capabilities accelerate exponentially, enterprise deployment rates are improving linearly. The gap is widening, not narrowing.
The Salesforce data — 12 agents per organization, 50% in silos — describes the exact problem AgentPMT's centralized architecture solves. Instead of 12 siloed agents with 12 different tool configurations and 12 different cost structures, AgentPMT provides one integration point through Dynamic MCP's hub-and-spoke architecture to the largest marketplace of AI tools and AI skills, with unified audit trails, budget controls, and cross-platform compatibility. Every workflow, every tool call, every dollar is tracked from one dashboard and one mobile app.
The Enterprise Push Is Happening Anyway
Despite the deployment gap, enterprises are pushing forward — and discovering the same friction points.
Amazon mandated 80% weekly AI coding tool usage across its engineering organization on February 17. Seventy percent of its developers have tried Kiro at least once. But the mandate backfired in one telling way: Amazon flagged OpenAI Codex as "Do Not Use," and 1,500 employees petitioned for access to Claude Code. The developer revolt over tool restrictions is the canary in the enterprise mine.
Infosys partnered with Anthropic the same day to build enterprise-grade AI agents for telecom, finance, and manufacturing. Under the deal, Infosys integrates Claude models into its Topaz AI platform to build agentic systems for heavily regulated industries. AI-related services already generate roughly $275 million for Infosys, representing 5.5% of total revenue. As Anthropic CEO Dario Amodei told TechCrunch: "There's a big gap between an AI model that works in a demo and one that works in a regulated industry." Infosys's experience in financial services, telecoms, and manufacturing helps bridge that gap.
OpenAI launched Frontier on February 5 as an enterprise agent management platform, with HP, Oracle, State Farm, and Uber as launch customers. Agents are "onboarded like employees" — with roles, permissions, and audit trails. Enterprise revenue now represents 40% of OpenAI's total, with a target of 50% by end of year.
Anthropic's Claude Cowork landed on Windows with full feature parity — file access, multi-step task execution, plugins, and MCP connectors — and enterprise plans became purchasable directly without a sales team. VentureBeat reported Microsoft is spending roughly $500 million annually on Anthropic, deploying Claude tools internally while simultaneously selling GitHub Copilot externally. Microsoft's CoreAI team has tested Claude Code across all code and repositories.
The SaaSpocalypse continues as backdrop. Software stocks shed $285 billion in the week following Cowork's announcement. The S&P 500 Software & Services Index dropped over 20% year-to-date. Hedge funds shorted $24 billion in software stocks. Goldman Sachs responded by analyzing which software companies have "moats" against AI disruption — a tacit acknowledgment that the disruption is real even if panic may be overblown.
Amazon's developer revolt over tool restrictions directly validates the cross-platform architecture AgentPMT was built on. AgentPMT's Dynamic MCP works identically across Claude Desktop, Claude Code, ChatGPT, Codex CLI, Gemini CLI, Cursor, VS Code, Windsurf, and Zed — one workflow, every platform. Developers don't petition when they aren't locked in. OpenAI Frontier's "agents onboarded like employees" approach aligns with AgentPMT's philosophy — budget controls, audit trails, tool whitelisting — but locks enterprises to OpenAI's ecosystem. AgentPMT achieves the same accountability structure without platform lock-in.
MCP: The Standard That Won — And the New Problem It Created
The Model Context Protocol has won the protocol standards war. The question is no longer "which protocol?" but "how do you manage thousands of MCP connections at enterprise scale?"
CData reports 8,620 MCP server implementations, with Gartner predicting 40% of enterprise applications will incorporate agents by year-end. The adoption milestones landed fast. Atlassian's Rovo MCP Server went generally available on February 4, making Jira and Confluence accessible from any MCP-compatible agent with domain allowlists for admin control. Amazon Ads launched an MCP server open beta. Microsoft previewed a Power Platform MCP server. Apple adopted MCP for Xcode 26.3 agent-IDE communication. VS Code 1.109, released February 5, became a multi-agent development platform running Claude, Codex, and Copilot agents in parallel. Vercel's AI SDK 6 added agent abstraction and MCP support for its 20-million-monthly-download TypeScript toolkit.
The governance signal was equally strong. MCP was donated to the Linux Foundation's Agentic AI Foundation, co-founded by Anthropic, Block, and OpenAI. Protocol neutrality achieved.
But MCP's victory is both the solution and the new problem. With 8,620 servers and adoption by Apple, Microsoft, Google, Amazon, and Atlassian, MCP is the de facto standard. "Standard" doesn't mean "manageable." Enterprises connecting to hundreds of MCP servers face the same complexity problem they had with APIs — except now the consumers are autonomous agents, not developers reviewing documentation.
This is AgentPMT's core value proposition. Dynamic MCP is the management layer between agents and the 8,620-server MCP ecosystem. Instead of connecting each agent to each server individually — creating the silo problem Salesforce's own data identified — Dynamic MCP provides one connection point that routes to the right tools on demand. The lightweight 5MB binary costs $0, auto-detects installed AI platforms, and the tool catalog updates every 30 minutes without any action from you. Remote tool fetching means agents access the largest marketplace of AI tools and AI skills without consuming a single context token until called. One install. Unlimited access. Zero bloat.
What This Means For You
The capability-deployment gap is the defining challenge in enterprise AI right now. Stop evaluating agent capabilities. They can use computers. Start evaluating deployment infrastructure. The framework determines whether your 72.5%-capable agent produces business value or joins the 40% Gartner says will be cancelled.
Cross-platform flexibility is non-negotiable. Amazon's developer revolt proved that locking teams to a single AI vendor creates resistance, not productivity. Your deployment infrastructure must work across Claude, ChatGPT, Codex, Gemini, and whatever ships next month.
Cost visibility must be built in from day one. Copilot's 3.3% adoption happened partly because enterprises couldn't see per-task cost impact clearly enough to justify $30 per seat per month across hundreds of thousands of users. AgentPMT's credit system — 100 credits equals $1, only charged on successful tool calls, failed calls automatically refunded — provides the per-transaction cost transparency that per-seat licensing fundamentally can't.
For vendors and tool developers: if your tool isn't accessible via MCP, you're invisible to 8,620 agent integrations and growing. The marketplace for AI tools is consolidating around hub-and-spoke architectures. Get listed where agents look first.
What to Watch
Sonnet 4.6's 72.5% OSWorld score is a benchmark. Watch for enterprise case studies showing real-world task completion rates and error patterns. If production accuracy tracks the benchmark, computer-use agents become viable for high-volume back-office workflows within months.
Amazon's developer tool mandate will play out publicly. The 1,500-employee petition for Claude Code is a leading indicator. If Amazon loosens tool restrictions, it validates cross-platform flexibility over vendor lock-in. If it doubles down, expect more developer friction stories from every enterprise trying the same approach.
Microsoft Copilot's 3.3% adoption rate is embarrassing for the most widely distributed AI product on the planet. If paid adoption doesn't reach 5-7% by mid-2026, it validates that enterprise AI needs a different delivery model than per-seat licensing.
And with 8,620 MCP servers and a known 36.7% SSRF vulnerability rate across analyzed implementations, enterprise MCP adoption will be bottlenecked by security until the Agentic AI Foundation and CoSAI deliver standardized security frameworks.
The agents learned to use computers this week. Sonnet 4.6 navigates spreadsheets. Fujitsu's platform compresses months into hours. Qwen 3.5 controls mobile applications. The capability question is settled. The deployment question is just getting started — and the companies that solve it with accountability, cost transparency, and cross-platform flexibility built in from the first deployment are the ones that turn AI capability into business value instead of another cancelled project.
Key Takeaways
- AI agents crossed the computer-use capability threshold in February 2026, but enterprise deployment rates remain stuck at 3-11% — the framework for accountability and cost transparency is now the bottleneck, not the model
- Microsoft Copilot's 3.3% paid adoption rate, despite 450 million potential users, demonstrates that per-seat licensing fails as a delivery model for AI agents
- MCP has won the protocol standards war with 8,620+ server implementations, but enterprises now face a new problem: managing thousands of MCP connections at scale requires a management layer like AgentPMT's Dynamic MCP
Sources
Anthropic releases Claude Sonnet 4.6 - CNBC
Introducing Claude Sonnet 4.6 - Anthropic
Fujitsu automates entire software development lifecycle - Fujitsu Global
Alibaba Unveils Major AI Model Upgrade Ahead of DeepSeek Release - Bloomberg
A new version of OpenAI's Codex is powered by a new dedicated chip - TechCrunch
Microsoft Finally Revealed How Many Paying Copilot Customers It Has - The Motley Fool
The State of AI in the Enterprise 2026 - Deloitte
Salesforce Connectivity Benchmark 11th Annual - Salesforce with Vanson Bourne and Deloitte Digital
Amazon Pushes Developers Toward Weekly AI Coding - Business Insider / Startup News
As AI jitters rattle IT stocks, Infosys partners with Anthropic to build enterprise-grade AI agents - TechCrunch
OpenAI launches a way for enterprises to build and manage AI agents - TechCrunch
2026: The Year for Enterprise-Ready MCP Adoption - CData Blog
Atlassian Rovo MCP Server is now GA - Atlassian Blog
VS Code 1.109 Deemed a Multi-Agent Development Platform - Visual Studio Magazine
AI SDK 6 - Vercel Blog
Donating the Model Context Protocol and establishing of the Agentic AI Foundation - Anthropic
AI fears pummel software stocks: Is it illogical panic or a SaaS apocalypse? - CNBC
Anthropic's Claude Cowork finally lands on Windows - VentureBeat
Goldman says these software stocks have moats that can thwart AI disruption - CNBC