Multi-Agent System

The strongest argument for multiple agents is context hygiene. A single agent handling research, code generation and review accumulates every intermediate artefact in one window, and quality degrades as the relevant instructions compete with thousands of tokens of history. Splitting the work gives each agent a small, focused context and a narrow tool catalogue, which improves both tool-selection accuracy and instruction adherence. A second, weaker argument is parallelism — independent subtasks can genuinely run at the same time.

The cost is coordination, and it is larger than it looks. Every hand-off is a lossy serialisation: whatever the sending agent knew that it did not write down is gone. Errors propagate silently because the receiving agent has no way to judge whether the input it was handed is sound. Debugging requires reconstructing a distributed trace across several conversations, and total token spend rises because context is duplicated across participants.

The misconception is that more agents means more capability. Adding agents adds interfaces, and interfaces are where systems fail. The reliable progression is to start with one agent and good tools, split only when a specific context or permission boundary forces it, and keep hand-off payloads explicit and structured rather than trusting free-form prose to carry state between them.