The shift in how knowledge workers relate to AI assistants has been gradual but unmistakable. For years, the dominant pattern was retrieval and advice: ask a question, get a summary, take the answer away and do something with it. The assistant informed; the human still did the work. That pattern made sense when AI could not reliably sustain multi-step tasks without constant correction. It is starting to look like a transitional phase.
From “told” to “done”
Most AI tools today still operate in what might be called “told” mode — they summarize, brief, and surface, but leave the actual production work to the person asking. The gap between “the AI told me the answer” and “the AI handed me the finished thing” is where most of the remaining friction lives. Closing that gap requires something AI assistants have rarely had: a persistent, shared environment where different AI roles can hand work to each other without a human shuttling context between windows.
The architectural move that makes this possible turns out to be surprisingly simple in concept: a single, account-scoped computer — one persistent browser, file system, and set of credentials — shared across all the AI agents on a team. When one agent saves a file, another can pick it up directly. No download, no re-upload, no re-briefing. The human is no longer the integration layer.
What changes when the environment is shared
Early multi-agent systems put coordination back on the user. If one AI role needed output from another, the user copied and pasted, reformatted, re-explained. The shared compute model removes that tax. Agents can pass work to each other asynchronously — one finishes a research brief, another picks it up and drafts an outreach email — and the human only re-enters at moments that genuinely require judgment or authorization.
This changes what good agent design looks like. The instinct when setting up an AI agent is to define a task: “summarize this thread,” “find three competitors.” But agents designed around tasks finish and go quiet. The more durable design gives each agent a continuing theme or domain to own — a landing page, a research pipeline, a scheduling function — rather than a one-off job. Theme-ownership is what makes an agent feel like a persistent team member rather than an on-demand API call.
The onboarding instinct to resist
There is a familiar failure mode in any tool that promises broad capability: it asks for broad access before it has demonstrated any value. The temptation for both product teams and users is to connect everything upfront — every inbox, every calendar, every data source — before testing whether the core job even works.
The more disciplined approach inverts this. Start with a single document or data source. Get the first run right. Add connectors only when the work actually reaches them. Broad access before demonstrated value is not a productivity accelerator; it is a risk surface with no confirmed benefit yet.
The same logic applies to autonomy. Reading, preparing, and organizing work can be handed over with relatively light oversight. Sending, publishing, purchasing, and deleting — anything with a consequence outside the system — stays behind human approval. The distinction is not about distrust; it is about matching the cost of an error to the level of oversight in place.
The security reality of a shared environment
A multi-agent setup running on a shared environment is not a collection of isolated security zones. All agents share the same signed-in sessions and file system. A bad instruction — or a poorly scoped agent that acts on the wrong source — has access to everything the team does. The roster of agents on a shared computer should be treated the way one would treat a team with shared keys to the office: trusted, deliberately composed, and not expanded casually.
This is not a reason to avoid the model. It is a reason to think carefully about which agents get added, and in what order, before the environment is woven into consequential workflows.
Where the practice is heading
The real question these systems raise is not “can AI do my job” but “what does my job look like when the gap between idea and finished artifact is measured in minutes rather than days.” The early advantage is not access to better AI — it is knowing which jobs to assign, how to define “done” well enough for an agent to reach it, and where to stay in the loop. Those are skills, and like most skills, they develop best by starting small, watching what breaks, and expanding scope from there.
Text summarized and optimized using Anthropic’s models and reviewed by a human.