Six of the largest AI labs on the planet shipped what is essentially the same product in 120 days, and not one of them seems sure their customers know what to do with it. Anthropic, Perplexity, Microsoft, OpenAI, Google, and Amazon all rolled out a desktop agent that reads local files, drives the browser, holds context across days, and produces finished outputs instead of suggestions. The overlap is hard to miss. The adoption story is messier.
Why Every Lab Shipped the Same Knowledge Worker Agent
According to The New Stack’s reporting, the timeline is tight enough to fit on a sticky note. Anthropic shipped Claude Cowork on January 12 on top of Claude Code’s agentic harness — three weeks later, investors knocked $285 billion off the SaaS index when the open-source plugin pack hit GitHub. Perplexity launched Computer on February 25, routing work across nineteen models with Claude Opus as the reasoner. Microsoft announced Copilot Cowork on March 9, built on Claude via its expanding partnership with Anthropic. OpenAI rebuilt the Codex desktop app on April 16 with computer use, ninety plugins, persistent memory, and scheduled automations. Google unveiled the Gemini Enterprise Agent Platform on April 22, complete with an inbox of long-running agents. Amazon launched Quick on April 28 with a personal knowledge graph and connectors for Google Workspace and Microsoft 365.
This is what category capture looks like when the underlying capability — a model good enough to wield tools autonomously for hours — finally crosses a threshold. None of these labs is being original. Each one watched developers fall in love with Claude Code and asked the same question Kate Jensen put to CNBC: why should this stay a developer tool? Six product orgs ran the same play because the play obviously works in one vertical, and the labs are betting it generalizes.
If you are an enterprise architect, this means the procurement decision is no longer about choosing a winner. It is about deciding which agent harness will sit on top of the productivity suite your employees already live in. The lab that owns the calendar, inbox, and document store wins by default, regardless of model benchmarks. The harness is becoming a commodity. Context is the moat.
The Developer-to-Knowledge-Worker Translation Problem
Here is the fact the press release decks gloss over: knowledge workers are not developers. Claude Code’s adoption curve assumed an audience that already lived in a terminal, understood file systems, and could read a stack trace. The Cowork pitch asks marketing managers, finance analysts, HR leads, and operations teams to do the thing engineers spent two years learning — delegate a multi-step task, supervise an agent doing the work, catch the moment it goes off the rails, and trust an output that was not produced one keystroke at a time.
The gap between the six labs is not technical. It is behavioral. The product that wins will be the one that fits the existing rhythm of how non-engineers work — approvals, audit trails, reviews — rather than the one with the most impressive demo. Microsoft has the rhythm. Google has the Workspace context graph. Amazon has the data graph. Anthropic has the harness. Perplexity has the orchestration. OpenAI has the brand and what the source reports as three million weekly developers it can drag forward.
Imagine you are running operations at a mid-market firm. Your finance team does not want a chatbot. They want something that picks up an invoice exception, traces it back through three systems, drafts a resolution email, and waits for sign-off before sending. That is closer to AI automation than to a conversational agent, and the labs that recognize the distinction will outsell the ones that keep shipping demos of agents booking flights. By Q4 2026, the marketing copy on every one of these products will quietly shift from “agent that works like a coworker” to “automation with judgment.”
What the Microsoft and PwC Numbers Actually Reveal
Microsoft’s April earnings call disclosed twenty million paid Copilot subscribers, up from fifteen million in January — a 33 percent jump in a single quarter, though still under 5 percent of the 450 million commercial Microsoft 365 base, as The New Stack notes. The question is whether those buyers turn seat licenses into daily habits. PwC’s April commitment to roll out Cowork and Claude Code to hundreds of thousands of professionals worldwide is the largest enterprise cowork deployment disclosed so far, and it is the single data point most worth watching.
Enterprise software has a long history of seat counts that never translate into usage. If PwC’s rollout produces a viral internal demo — a partner who closes an engagement two days faster because an agent assembled the deliverables overnight — the procurement curve for every other firm compresses from quarters to weeks. If it does not, this looks more like the slow grind of any other enterprise platform: tied to approval workflows, change management timelines, and middle-manager incentives.
For teams evaluating whether to wait for the SaaS vendors to ship Cowork-style features or build something tailored now, the calculus is the same one custom AI versus off-the-shelf software has always faced. The off-the-shelf agents will be good enough for generic tasks within a year. They will not understand the eighty proprietary fields in your CRM. Most firms will end up running both — a horizontal agent from one of the six labs for general work, and purpose-built agents for the workflows that actually move revenue.
FAQ
Q: What is a knowledge worker agent? A: It is a desktop AI agent that reads local files, drives a browser, retains context across multiple work sessions, and delivers finished outputs instead of suggestions. The six examples shipped between January and April include Anthropic’s Claude Cowork, Microsoft’s Copilot Cowork, OpenAI’s rebuilt Codex desktop, Google’s Gemini Enterprise Agent Platform, Amazon’s Quick, and Perplexity’s Computer.
Q: Why did six labs converge on the same product so quickly? A: Claude Code proved that agentic harnesses on frontier models could ship real work for developers. According to The New Stack, every lab watching that adoption curve concluded the same pattern would extend to non-engineers, so each ran the same product play in parallel rather than risk being last to market.
Q: Is enterprise adoption actually happening? A: The early signals are mixed. Microsoft disclosed twenty million paid Copilot subscribers in April, up from fifteen million in January, but that is still under 5 percent of its 450 million commercial Microsoft 365 base. PwC’s commitment to deploy Cowork and Claude Code to hundreds of thousands of professionals is the largest disclosed rollout so far.
Key Takeaways
- Treat the agent harness as commoditized — choose your vendor based on which productivity suite owns your team’s context, not on model benchmarks.
- Watch the PwC rollout closely; a single viral internal case study from that deployment will reset enterprise procurement timelines for every competitor.
- Plan for hybrid agent strategies — horizontal Cowork-style agents for generic work and domain-specific agents for the workflows that matter to revenue.
- Invest in change management before invest in seats; the 5 percent Copilot penetration number suggests license growth is outpacing actual habit formation.
- Expect vendor marketing to pivot from “agent as coworker” to “automation with judgment” as labs realize knowledge workers want supervised delegation, not an AI peer.