What comes next

Predictions built on what exists today, sorted by how strong the evidence is

Why this page is different

Other pages describe what exists now. This page holds predictions, not announcements. No vendor has promised these features. They are our view of where today's tools lead. The agent now has a disposable computer of its own, run from a server. So a session can be timed, started, moved, sent notes, checked and paid for, like a server job. Most of what follows comes from that.

Signaled: the documents already point here

1. Projects for teams. [Signaled] Today "a project belongs to one user."¹ Anthropic's launch post says projects for "Team and Enterprise plans come after that."² The guess is shared projects with roles. One person sets the brief. Others add work. A manager approves.

2. Workers that wake on business events. [Signaled] Cloud scheduled tasks "run in the cloud on a schedule, on API calls, or on GitHub events."³ The Tour records that ChatGPT's scheduled tasks can start on Gmail, Slack, GitHub and plugin events. The guess is built-in triggers from many more business systems. A new invoice in Brightline's inbox could start a run.

3. Workers with their own identity. [Signaled] In Slack channels, Claude Tag "acts with its own service accounts, rather than as a specific user."⁴ A self-hosted session carries a signed token. It says whether a person or "your organization's service identity" created it.⁵ The guess is a worker inside the company's identity system, with its own login, roles and access that can be cut off.

Assembling: the pieces exist and need joining up

4. Approval and policy as a platform service. [Assembling] Today an approval is a prompt inside one thread. A limit typed to a coordinator is "not enforced settings."¹ The guess is one approval queue across all your workers. A second person signs anything over an amount. Each worker has a spending limit. The business record and its rules stay with the company. Brightline's payment connector still checks again.

5. Worker templates you can share and hire. [Assembling] A setup, plugins, instructions and a schedule now describe a role.⁶,¹ The guess is packaged roles, such as an "accounts payable worker". They would be shared the way plugins are shared today.

6. Testing workers like software. [Assembling] Setup snapshots give repeatable starting points. Anthropic "stores the session transcript."⁶,⁷ The guess is replaying a run against a changed brief, and staging a worker before it goes live. Brightline could test a new brief on last Friday's invoices first.

7. Desks that work old software. [Assembling] The Tour documents a ChatGPT cloud browser. It also documents computer use in Claude and ChatGPT. The guess is screen control on the cloud desk. A worker could then use old portals that have no API.

Speculative: further out

8. Workers inside your own network and country. [Speculative] Session content itself may come to stay inside the company or the country. For banks and public bodies, that may decide whether AI Workers can be used at all. Today's facts point part of the way. [Vendor-documented] Self-hosted runs cloud sessions on your own machines. But "session content goes to Anthropic in both cases."⁷ Claude Managed Agents lists "Sandboxes on infrastructure you control for compliance or data-residency requirements."⁸

9. Workers that talk across companies. [Speculative] Your own sessions can message each other now. A message "can't approve anything."⁹ Add verified identity across companies, and a buyer's worker could agree a delivery date with a supplier's worker.

10. Paying for outcomes, not seats. [Speculative] A project now reports "token use by thread and by model."¹ Pricing per finished job, such as per reconciled invoice, becomes possible.

11. The phone as the manager's desk. [Speculative] You can now steer project threads "from your phone."¹ The human job may become a stream of choices: say yes, change course, or help a stuck worker.

Putting it together: one larger prediction

12. Independent AI Workers in the cloud, measured as Digital FTEs, that talk to each other. Its four parts stand on different evidence:

• Deployed in the cloud. [Vendor-documented] Cloud threads "keep going after you close your laptop."¹ Self-hosted runners run the same sessions on a company's machines.⁷ [Speculative] The worker will likely stay a series of short sessions. Its continuity lives in memory, files and the business record.
• Independent in doing the work, not in deciding what it may do. [Vendor-documented] Today, workers start on events, retry and finish alone. Claude Tag acts under its own accounts.⁴ [Speculative] For money and legal commitments, the rule that authority stays outside the worker will likely tighten.
• Measured as Digital FTEs. [Speculative] FTE stands for full-time equivalent, a way to count work in units of one full-time person. This is our framing. A worker has a role, an identity, a trigger, a budget and an output you can check. But one worker can run many threads at once, or sit idle. So a Digital FTE is a unit of delivered work, not a headcount number.
• Able to talk to each other. [Vendor-documented] Today, within your own sessions and agent teams.⁹,¹⁰ [Assembling] within one company, once team projects, worker identity and shared environments join up. [Speculative] across companies and vendors. Open standards are on the way. One is the Agent2Agent protocol, "an open protocol created by Google for secure agent-to-agent communication and collaboration."¹¹ The hard part is trust.

AI Workers will run in the cloud as persistent, identified workers, measured as Digital FTEs. They will act on their own schedules and events. They will coordinate through messages. Messages will carry information, not authority. Each worker's permissions, records and approvals will stay with the business that deploys it.

The last sentence keeps the prediction safe. If messages could authorize actions, one worker could manipulate another. A supplier's worker could send Brightline a corrected invoice. Only Brightline's own approval could release the payment.

What does not change

Every prediction still depends on three lines:

• The sandbox limits reach but grants no permission (Security boundaries).
• Lasting state lives outside the sandbox (The sandbox and its state).
• A person or a system you control decides what may happen (From personal worker to business worker).

[Inferred] The platforms will take over more of the plumbing: identity, rules, triggers and tests. The design work then moves up a level: writing the brief, choosing the environment, setting authority and owning the record. That is the Forward Deployed Engineer's job.

Rule to remember: Features change. The pattern does not. Check what is a signal and what is a guess. Keep authority in the service that acts.

Next, after the appendices, see the first case study, Case 1: Nisha's AI Team. It shows these ideas at work in one real evening.