The idea of one enterprise agent is appealing. Give it access to everything, teach it how the organisation works and let people ask for what they need through a single interface.
I understand the attraction. I am also increasingly wary of it.
Complex work is rarely performed by one person with unlimited authority. We separate responsibilities because different tasks require different expertise, controls, and permissions. We use independent checks because diverse failure modes improve assurance. We escalate because no individual should decide everything.
Agentic architecture should learn from that.
A multi-agent design can separate the work into defined responsibilities. One agent may interpret the request. Another retrieves authoritative information. Another checks policy or regulation. Another proposes an action. A separate verifier tests the output, and a human remains accountable for consequential decisions. An orchestrator coordinates the flow without every component receiving unrestricted access.
This is not about creating a cast of digital characters. It is about engineering separation of duties and making each component easier to test, monitor, and improve.
Specialisation also makes performance more visible. If a result is poor, we can identify whether the failure came from retrieval, reasoning, policy interpretation, execution, or orchestration. A single generalist agent can hide those distinctions behind a fluent answer.
There are costs. Multi-agent systems introduce complexity, latency, and additional model calls. More components create more interactions to secure and monitor. The answer is not to use multiple agents everywhere. The architecture needs to reflect the complexity, value, and risk of the work.
This is where process design matters. We need to understand the outcome, break the work into meaningful stages and decide where judgement, validation, and authority belong. Only then can we decide whether one agent, several agents, deterministic automation or a person is the right choice.
I have moved away from asking how powerful an agent can become. I am more interested in how safely a system of people and agents can work together.
The best agentic architecture may look less like a digital super-employee and more like a well-designed team: specialised roles, limited authority, clear hand-offs, independent challenge and one accountable human who owns the outcome.