Imagine a status meeting where everyone uses AI. Research is faster, the presentation is ready, and documents have been summarized. Yet the first question is familiar: who is waiting for whom, and what has to happen before the work is finished? For us, that is a useful starting point for a conversation about agent teams.
Capability belongs to the worker. Responsibility spans the process.
An individual assistant can be excellent at a defined assignment: writing, analysis, or research. In an organization, however, its output usually becomes someone else’s input. Research must inform a brief, the brief must become execution, and execution must be reviewed. Each handoff contains a decision about whether the work is good enough to continue.
When those handoffs are undefined, the manager becomes the coordination layer. They gather outputs, restore missing context, explain priorities again, and keep an earlier mistake from moving downstream. A useful discussion of business value therefore includes the work the manager still does around all those intelligent outputs.
A team starts with a clear working agreement
Our starting point for designing a digital team is the intended result. What must be ready? Who can make a decision? Which sources may be used? What must an output contain so the next stage can rely on it? We define roles and tools after answering those questions.
Consider preparing a legal working file as an illustrative example. The assignment can branch into document collection, chronology, source cross-checking, and drafting. A process manager needs to recognize when a missing document blocks a conclusion, return an unsupported claim, and show the lawyer what still requires judgment. The team’s quality depends on holding those relationships together.
Hierarchy should earn its place
In AOS, an AI CEO translates organizational objectives into work plans and delegates to agent managers. Those managers direct specialist subagents, review outputs, and report progress. This is a management role inside the system. Human leadership defines policy, authority, and the decisions it retains.
Every task does not need several managers and agents. A clear, simple action may need one tool or an automation. We look for a reason behind each management layer: specialization, parallel work, separated permissions, or independent review. If a layer merely creates more handoffs, the design should become simpler.
The technical pattern is established beyond GOBOOST: Anthropic describes an orchestrator that delegates to workers and combines their results, while recommending the simplest adequate solution first. This supports an architectural choice; it does not establish the performance of any particular product. Building effective agents — Anthropic
Measure the work around the output
Alongside output volume, we suggest measuring acceptance without revision, time spent waiting, context that people had to restore, and behavior when a source was missing. One process may prioritize exception-resolution time. Another may prioritize source coverage or the ability to resume an interrupted assignment.
A useful first implementation is bounded, has an owner who understands it, and has completion criteria that can be checked. Map the handoffs, choose a small team, define quality gates, and expand in response to findings. The AI discussion then becomes a familiar management question: how do we give a team responsibility and understand what it does with it?
Choose one process and ask who owns the responsibility between its stages. That is where a collection of AI tools starts becoming a working team.
Explore AI teams in action