Multi-agent Systems

A multi-agent system is not “a few prompts in a loop”. It is a distributed workflow where each participant has its own responsibility, typed message exchange and explicit coordination. Without that, a set of agents becomes an unmanageable chain whose behavior can’t be reproduced twice.

Roles and contracts

Each agent: one task, a constrained tool set, a fixed input/output format. Inter-agent contracts are pinned, and that is what makes the system testable: an agent can be run in isolation with a known input and its output checked without bringing up the whole workflow. An agent without a contract can neither be tested nor replaced without breaking the rest.

Coordination

Ordering, branching, retries and termination are run by an orchestrator, not by agents “negotiating among themselves”. Workflow state lives outside the agents — in the workflow, not in a single step’s context. That is what separates an engineered system from an emergent one: in the first you can say at which step and why a decision was made, in the second you can’t.

Fault tolerance

Timeouts, fallback to a simpler strategy, humans as the terminal handler. One agent’s failure moves the workflow to a predefined state instead of collapsing it. A looping agent stops by rule and escalates, rather than burning resources until timeout.

Where the line is

More agents means more latency and coordination cost. Each split is justified by what it buys: risk isolation, testability, the ability to replace a part. Decomposition driven by “make it multi-agent” adds complexity without adding value. The right question is not “how many agents to split into” but “which split makes the risk controllable”. Often two or three roles with strict contracts are enough where a swarm seemed necessary.