Anthropic's Claude Cowork launched in January 2026 as a research preview. Within weeks, it had contributed to a trillion-dollar selloff in software stocks. Thomson Reuters dropped 15.83%. LegalZoom fell 19.68%. The market decided AI agents were coming for enterprise software.
The product is real. The panic? Premature.
Cowork represents something genuinely new in how AI gets packaged for mainstream users, but the selloff reflects market sentiment racing ahead of what the technology can actually do today.
Claude Code without the command line
Simon Willison tested the product on launch day and nailed the description: Cowork is Claude Code wrapped in a less intimidating interface. The underlying capabilities are the same agentic architecture that powers Claude Code (file system access, autonomous task planning, multi-step execution). The difference is packaging. Where Claude Code targets developers comfortable with terminals, Cowork targets everyone else: analysts, researchers, marketers, lawyers. You point Claude at a folder, describe what you want done, and it figures out how to do it.
Planning is built in. Claude creates a task list, works through it systematically, and prompts you for approval on significant actions. You can queue multiple tasks for parallel processing. The model can perform dozens of steps autonomously before checking in.
How autonomous, exactly? Willison pointed Cowork at his drafts folder, and Claude performed 44 web searches autonomously to verify the publication status of 46 draft files.
No hand-holding required.
Plugins are where enterprise software gets nervous
The real escalation came on January 30, when Anthropic launched plugins for Cowork. This is where the enterprise software threat becomes concrete. Plugins bundle four components: skills, connectors, slash commands, and sub-agents. Skills define capabilities. Connectors link to external data sources. Slash commands provide shortcuts for common workflows. Sub-agents are specialized Claude instances that handle specific tasks within a larger workflow.
Anthropic open-sourced 11 production plugins at launch covering legal, finance, sales, marketing, customer support, product management, data analysis, and biology research. These are not demos. They are working systems designed for enterprise customization. The legal plugin can review documents and flag risks. The finance plugin builds models and tracks metrics. The sales plugin connects to CRMs and knowledge bases. Each plugin tells Claude how you like work done, which tools to pull from, and how to handle specific workflows.
The architecture is file-based, making plugins easy to build, edit, and share. Organization-wide sharing and private marketplaces are coming. Anthropic's PM Matt Piccolella emphasized that plugins are built for enterprise customization, positioning Cowork as a potential alternative to traditional point-solution software.
There is a significant caveat that gets buried in the capability hype. Cowork runs in a containerized sandbox using Apple's VZVirtualMachine framework, which provides meaningful isolation; if Claude tries to do something malicious, it stays constrained to the sandbox environment. But the bigger concern is prompt injection. Cowork gives Claude access to files you designate. If one of those files contains adversarial instructions, Claude might follow them. Anthropic acknowledges this directly: prompt injection defenses are listed as an "ongoing development area."
Willison flags this as a real problem: telling regular users to "monitor for prompt injection" is unreasonable. Most users do not know what prompt injection is. He predicts a high-profile incident before users take the threat seriously.
Our read: This is not a dealbreaker for controlled environments where you trust your inputs. For knowledge workers processing external documents at scale, it is an unsolved problem.
The trillion-dollar question
The selloff was not about Cowork alone. It was Cowork plus Opus 4.6's agent teams feature, plus the 1-million-token context window, plus the demonstrated capability to perform tasks that generate significant SaaS revenue today.
Fortune's reporting captured the bear case cleanly:
Claude is demonstrably superior for "financial research tasks such as screening, due diligence data gathering, and market-intelligence synthesis".
That describes core workflows at FactSet, Thomson Reuters, and LegalZoom. But Gartner's analysts offered the counterpoint (via the same Fortune piece): these tools are "potential disrupters for task-level knowledge work but are not a replacement for SaaS applications managing critical business operations." The gap between "can automate a task" and "has replaced mission-critical enterprise software" is where the real debate lives. Wedbush analysts argued the panic is premature. Ingrained organizational workflows resist overnight transitions. We covered this tension directly: the market is pricing in contradictory outcomes, betting AI is simultaneously too weak to justify infrastructure spending and too strong for SaaS companies to survive.
Cowork is a real product with real capabilities. It is not vaporware. When Simon Willison says it performed 44 autonomous web searches on his drafts folder, that is not a benchmark; that is a Tuesday. The question is adoption.
Enterprise software is sticky. Workflows are embedded in organizational processes. The tools that got hit hardest in the selloff (legal research, financial data, document review) are the ones where capability overlap is clearest. For companies selling those services, the competition is real and arriving now. For CRM vendors and broader SaaS, the threat is more indirect. Claude needs your data even if it replaces your interface. The transition from "software humans use" to "infrastructure Claude talks to" will not happen overnight.
Our read: The trillion-dollar bet is that this transition happens fast. The next two quarters of enterprise renewal data will tell us whether the market got ahead of itself or saw something real.