In February 2026, METR (Model Evaluation and Threat Research), the group behind the most-quoted study on AI coding productivity, said it was changing its experiment design. The reason was its own participants. Between 30 and 50 percent of its developers said they were skipping some tasks rather than do them without AI. I think that refusal says more about Claude Code and tools like it than any speed benchmark. People have already moved their work somewhere new, and they don't want to move it back.
Claude Code is Anthropic's agentic coding tool, and it now runs in the terminal, inside IDEs (integrated development environments), in a desktop app with a built-in browser, and on the web. It pulls local files, Git history, browser sessions, and project memory into one AI-first environment, so the agent operates your other tools from a single place. For product and marketing leaders, that makes an AI agent one of your most important users, and it will favor the products it can call reliably through APIs (application programming interfaces) and Model Context Protocol (MCP) connectors.
Why did SaaS make the browser tab problem worse?
The dominant software-as-a-service (SaaS) strategy of the 2010s rewarded products for capturing time-in-app. Every vendor built toward becoming the destination where your workday happens. The result is dozens of "platforms" competing for the center of your screen, and very few of them share context with each other.
Technical marketers feel this more than most, because we touch more tools than almost any other team. A single campaign runs through a customer relationship management (CRM) system, an analytics platform, a content management system, a project tracker, and usually a few AI chat windows that know nothing about each other. When I first wrote this piece in May, my laptop fan started straining somewhere around tab twenty-two, and three of those tabs were the same Google Doc. We've gotten so used to the friction that we just call it work.
The first wave of AI tried to fix fragmentation with a plugin for every app. Gmail got a sidebar, Notion got a slash command, and Figma got an AI panel. Each one helps in isolation. The architecture still brings the agent to the app, though, so the agent only sees what that one app shows it. You end up with a smarter participant trapped inside the same fragmented room.
How does Claude Code change where work happens?
Claude Code reverses that architecture by pulling the work into one environment. From there it can see local files, Git history, running terminal processes, browser state, and the project's accumulated instructions.
Most coverage of Claude Code focuses on what it can do, like writing code, running tests, and driving a browser. I care more about what it remembers. Every session starts by reading the project's CLAUDE.md file, a plain-text brief with your standards, decisions, and constraints, and Claude Code adds its own auto memory as it learns how the project works. You stop re-explaining your project every time you open a new window, so your attention goes to the problem itself.
With its Chrome integration, Claude Code opens tabs in your own browser, shares its logged-in state, and reads console errors and page structure (the DOM, or Document Object Model) directly. In July 2026, Anthropic added a built-in browser to the Claude Code desktop app, so the agent can read and click through docs, designs, or any other site, with safety classifiers reviewing its actions on external pages. In practice, the coding environment ends up running your other apps for you.
It also runs in more places than it did when I first wrote this. The same engine runs in the terminal, VS Code, JetBrains, a desktop app, and the web, runs scheduled routines in the cloud, takes bug reports from Slack, and packages repeatable workflows as shareable skills. Google is chasing the same idea with Antigravity's agent-first coding model. If you want to see the pattern applied outside engineering, I walked through building a Marketing Budget & ROI Advisor with Claude Code step by step.
What does the research say about AI coding productivity?
AI speeds up single tasks a lot and whole workflows much less. The difference comes from the work around each task. A 2023 controlled experiment by researchers at Microsoft Research, GitHub, and MIT found developers using GitHub Copilot finished a self-contained task 55.8 percent faster than a control group. Two years later, METR ran a randomized controlled trial with 16 experienced open-source developers working in their own mature repositories. With AI tools allowed, they took 19 percent longer to finish real issues, even though they believed AI had sped them up by about 20 percent.
METR's February 2026 follow-up widened the study to 57 developers and more than 800 tasks. Its rough estimates now lean toward a speedup, with confidence intervals wide enough to still include a slowdown. METR says developers are likely faster now than in early 2025, and it calls its own data weak evidence of how much.
My takeaway is that task-level gains only reach the organization when the environment absorbs the work around the task. Written as an equation, total time on any piece of work is the task itself, plus the handoffs between tools, plus the review that follows.
Think of it like a relay race. A faster runner helps, but the race is often lost at the baton exchange. AI point tools shrink Ttask, and they often grow the other two terms, because someone has to carry context between windows and check a bigger pile of output. The hypothetical numbers below apply the 55.8 percent task speedup to a two-hour job.
Both AI rows get the same task speedup. The AI-first environment still finishes 31 percent faster than the point tool, because it owns the handoffs and tightens the review loop. The point tool, in this example, ends up slower than working without AI at all. Claude Code goes after those last two terms by keeping writing, testing, browser checks, and deployment in one session.
How does Claude Code change product strategy?
Claude Code changes who your primary user is. If you own platform strategy as a product leader or chief marketing officer (CMO), that shift should reshape how you think about integration and differentiation.
The old model assumed a person navigating a graphical interface, so user experience (UX), visual design, and feature surface area decided who won. The emerging model puts an AI agent in the driver's seat, with a person supervising from an environment like Claude Code. These two models ask for almost opposite design priorities.
An agent judges your product on three things: whether your API is reliable, whether your outputs are machine-readable, and whether your error states are specific enough to recover from programmatically. If your product is hard for an agent to operate, the agent will route around it toward whatever is easier to call, however much your human users love the interface.
The table below maps that difference.
I think of the strategic result as the shift from destination to capability node. A destination tries to be where work happens. A capability node does one thing extremely well and exposes that capability so AI agents can discover it, call it, and compose it with other nodes. The products that win in an agent-driven market will be the ones an orchestration layer can use correctly every time. I made a similar argument about agentic AI and legacy marketing automation, and the same logic applies to SaaS product design.
Why does MCP matter for agent-ready products?
MCP is the open standard that lets AI tools connect to external data and software through one common interface. It works like a universal power adapter. You build one connector, and every compatible AI client can plug into your product. Anthropic released MCP in November 2024, and on December 9, 2025, it donated MCP to the new Agentic AI Foundation under the Linux Foundation, co-founded with Block and OpenAI and backed by Google, Microsoft, AWS, Cloudflare, and Bloomberg.
At the time of the donation, the ecosystem counted more than 10,000 active public MCP servers and 97 million monthly software development kit (SDK) downloads. ChatGPT, Cursor, Gemini, Microsoft Copilot, and Visual Studio Code all support it. That neutral home matters for the business case. An MCP server you ship today works with Claude Code and with whatever other agents your buyers run, so you aren't betting the build on a single AI vendor. It's the same question behind the Google Antigravity vs. n8n debate and of Amazon's push into AI agents.
What should product and marketing leaders do before the next planning cycle?
You'll see this shift first in roadmap discussions and integration requests, then in buyer questions your sales team can't answer yet. Three moves will get you ahead of it.
Audit your product through an agent's eyes. Open your API documentation, error messages, and onboarding flow, and ask whether an AI agent could get through them without a person translating each step. Any step that needs a human interpreter is your most urgent product priority.
Ask your engineering lead how Claude Code would use your product. The concrete version of that question is whether you have an MCP server, and what it would take to ship one. If nobody on the team knows the answer, that conversation is overdue.
Rewrite one buyer-facing asset for an agent reader. Pick a one-pager, a product page, or a capability slide, and rewrite it as if an AI agent were deciding whether your product belongs in an automated workflow. The exercise exposes assumptions in your value proposition that human-first messaging never had to defend.
All three fit in an afternoon, as long as you can get the right people in one room.
Will AI agents replace the browser tab?
The tab row will probably get longer, since agents open and close tabs faster than any person can. You just won't be the one managing it. Environments like Claude Code will run it for you while you work one level up. So the question for your team is whether your product and your messaging show up inside the AI-first environment that's doing the driving.
If this connects to something you're trying to solve, book a complimentary consulting session. No pitch, just perspective.
Frequently asked questions
What is Claude Code?
Claude Code is Anthropic's agentic coding tool. It reads your codebase, edits files, runs commands, and connects to other tools, and it runs in the terminal, in VS Code and JetBrains IDEs, in the Claude desktop app, and on the web. Its Chrome integration and the desktop app's built-in browser also let it operate websites and web apps directly.
How is Claude Code different from a regular AI chat assistant?
A chat assistant mostly sees what you paste into it. Claude Code works inside your project: it reads local files and Git history, runs terminal commands, operates a browser, and starts each session from a CLAUDE.md file plus its own auto memory, so it keeps project context from one session to the next.
Why should marketers and product leaders care about Claude Code?
Claude Code shows where work orchestration is heading. When one AI environment can pull data from SaaS tools, drive browser workflows, and hold context across long projects, the agent becomes a primary user of your product. That changes integration priorities, product messaging, and how buyers evaluate software.
What does "agent-ready" mean for a SaaS product?
An agent-ready product exposes its capabilities through reliable APIs and MCP connectors, returns machine-readable outputs, and surfaces specific error states an AI agent can interpret and recover from. The practical test: could Claude Code complete a core task in your product without a person translating each step?
What is the Model Context Protocol (MCP), and who governs it?
MCP is an open standard that lets AI tools connect to external data and software through one common interface. Anthropic released it in November 2024 and donated it to the Agentic AI Foundation, a Linux Foundation fund co-founded with Block and OpenAI, on December 9, 2025. Google, Microsoft, AWS, Cloudflare, and Bloomberg also back the foundation.
Do AI coding tools actually make developers more productive?
On isolated tasks, yes: a 2023 study found developers finished a self-contained task 55.8% faster with GitHub Copilot. On real work in mature codebases, the picture is mixed. METR measured a 19% slowdown in early 2025, and its February 2026 follow-up points toward a modest speedup that it calls weak evidence. The gains depend on how much of the surrounding workflow the tool absorbs.
What is a capability node?
A capability node is a product that does one thing extremely well and exposes that capability so AI agents can discover it, call it, and combine it with other tools. It contrasts with a destination product, which tries to be the place where the whole workday happens.
Can Claude Code use my browser and the apps I'm logged into?
Yes. With the Claude in Chrome extension connected, Claude Code opens tabs in your browser, shares its login state, and reads console errors and page structure. It pauses on login pages and CAPTCHAs for you to handle, and site-level permissions control where it can click and type.
