Multi-agent decision systems: bu ne anlama geliyor
Multi-agent architecture is fashionable, and most implementations of it would work identically as one model with a longer prompt. There is one case where the structure earns its complexity.
That case is genuine objective conflict. Purchasing wants inventory low. Production wants material available. Sales wants everything in stock. These are not misunderstandings to be resolved by better communication — they are real trade-offs with real owners.
Why one model handles this badly
A single model asked to balance all three produces an answer that reflects how the objectives were weighted in its instructions. The weighting is invisible, and the trade-off it made cannot be inspected.
Separate agents, each with one objective, produce a disagreement. The disagreement is the useful artefact: it shows what was given up and how much.
What this requires
A resolution mechanism — a policy that says who wins under which conditions, or an escalation to a person. Without it, multi-agent systems produce deadlock dressed up as deliberation.
The test for genuine conflict
Ask whether the two objectives could both be satisfied by a better answer, or whether one has to give.
A forecast that is wrong hurts purchasing and production alike; fixing it helps both, so there is no conflict — just a shared problem. Inventory level is different: every unit that reduces purchasing’s working capital is a unit production cannot draw on. No amount of accuracy dissolves that.
Only the second case justifies separate agents. The first is one problem being described by two departments, and giving it two agents produces theatre.
What the disagreement is worth
The artefact people underestimate. When purchasing’s agent argues for 40 units and production’s argues for 65, the record contains the reason for each — carrying cost against line-stop risk, with the numbers each side used.
That is the trade-off, written down, at the moment it was made. Most organisations have never had this. The decision was taken in a meeting, the reasoning stayed in the room, and six months later nobody can reconstruct why the number was 52.
Even where the resolution is a person rather than a policy, the disagreement makes their job possible: they are choosing between two stated positions rather than producing a number from instinct.
Resolution mechanisms, in order of preference
A policy is best where one exists: below this stock level production wins, above it purchasing does. It is inspectable, it is fast, and it can be argued with.
A weighted objective is next, and it is honest only if the weights are visible. A hidden weighting is the single-model failure with extra machinery around it.
Escalation to a person is the fallback, and it has to be rationed. A system that escalates twenty times a day will have its escalations ignored within a week, which is worse than not escalating at all.
What deadlock looks like in production
Not two agents arguing forever — that is the demo failure. The real one is subtler: a resolution policy that resolves cleanly on paper and always favours the same side, because the condition that would favour the other never occurs in practice.
Purchasing wins every time, production learns the system is not for them, and the multi-agent architecture becomes an expensive way to run one agent. Check the resolution log for balance before concluding the design works.
When one model is the right answer
Most of the time. If the objectives are aligned, if the trade-off is already settled by policy, or if the decision is small enough that the reasoning does not need to be inspectable, one model with a clear prompt is simpler and easier to operate.
Multi-agent is a structure for making conflict visible. Where there is no conflict, it adds latency and failure modes in exchange for nothing.