A personal team of specialists at one click. Claude Projects now acts as a manager, assembling subordinate agents tailored to the task

Most users still use LLMs as a lone assistant in a single chat window. However, Ethan Mollick pointed out that Claude Projects has quietly shifted to a different model. The main interface no longer represents one large language model trying to solve everything itself. Instead, you communicate with a main 'orchestrator'.

This main agent breaks down the task and dynamically spawns specialized subordinate agents in the background. It thus creates a temporary ad-hoc organization. In doing so, it combines expensive models for complex decision-making and cheap, fast agents for routine data processing. And all this without the user having to manually manage this process.

Costs are optimized in the background

For managers and founders, this is a key shift in how AI is paid for. Until now, the decision of which model to use (fast and cheap vs. slow and expensive) was up to the user before starting work. The orchestrator removes this dilemma.

The task is completed at optimal costs because the expensive intelligent model is used only to plan the work and for final synthesis, while the 'dirty work' in the intermediate steps is done by cheaper instructional variants. This approach significantly reduces the cost of solving a complex problem and at the same time increases speed, as independent subtasks can run in parallel.

Hidden vendor lock-in and dependence

As the system takes control over which models and tools are used at which stage, the user loses sight of the process. They become dependent on Anthropic correctly estimating which part of the task is trivial and which requires deep analysis.

If the orchestrator incorrectly assigns a complex subtask to a dumber agent, the overall result fails without it being immediately clear why. Users thus exchange transparency for convenience and gradually lose the ability to audit the system or easily migrate elsewhere because the entire organizational logic lies opaquely on the provider's side.

The future is a management dashboard

This trend shows that the future of the interface is not a better chat, but a better management dashboard. The next step will be the ability of these orchestrators to not only task virtual agents but also assign work to human employees in their real systems (Jira, Slack).

Ability to recognize bad output

We will be watching how these agents handle the mistakes of their subordinates. The proof that the system truly works as a team will be a situation where the orchestrator can recognize the poor output of a cheap agent, cancel its work, and assign it to a smarter, more expensive instance without human intervention.

Lilith's verdict

The days of prompting one omniscient window are ending. The future of the interface is hiring an agency manager, throwing a budget on their desk, and letting them sort it out with their robotic minions.