The simplest architecture that fits
Six architecture patterns for AI agents compared. Why traceability usually points law firms toward sequential workflows.
An agent architecture describes how many AI agents (programs that carry out several work steps on their own) work on a task and how they coordinate. Six patterns are common: single agent, hierarchical multi-agent, collaborative multi-agent, sequential workflow, parallel workflow and evaluator-optimizer. For law firms the sequential workflow is usually the right choice, because it is the only one that keeps every intermediate step traceable and ready for sign-off.
The most common mistaken assumption when building AI agents is that more agents produce better results. The real skill lies elsewhere: choosing the simplest architecture that meets the requirement.
Which architecture patterns exist for AI agents?
Alongside its well-known online courses, Anthropic also publishes written guides. One of them describes architecture patterns for agent systems and, more usefully, when each pattern fits. The overview is worth reading because it shows every pattern together with its price.
| Pattern | Fits when | Price |
|---|---|---|
| Single agent | the path to the result is unclear in advance | unusable where the first answer has to be right |
| Multi-agent, hierarchical | moderate need for control, flexibility under supervision | coordination overhead |
| Multi-agent, collaborative | low need for control, open-ended research | behaviour that is hard to predict |
| Sequential workflow | clear dependencies, sign-off chains, compliance | speed traded for accuracy |
| Parallel workflow | speed, several perspectives | unsuitable when steps build on each other |
| Evaluator-optimizer | clear evaluation criteria, rework pays off | loops cost time and tokens |
Single agent. One agent works freely toward a task. Useful when the path to the result is unclear in advance and nobody can say how many steps it will take. Avoid it wherever one correct answer matters and has to be right on the first attempt.
Multi-agent, hierarchical. A controlling agent breaks the task down, hands subtasks to specialised subagents that it treats as tools, and merges the results. It fits a moderate need for control, meaning situations where flexibility is welcome but supervised.
Multi-agent, collaborative. Agents coordinate directly among themselves, with no central control. This fits a low need for control and open-ended questions such as research. The price is behaviour that is harder to predict.
Sequential workflow. A fixed sequence, one step after another. It fits clear dependencies and anywhere traceability counts, such as sign-off chains or compliance. You trade speed for accuracy, because each individual step becomes simpler.
Parallel workflow. Independent subtasks run at the same time. This fits when speed counts or when several perspectives raise quality, for example when one model answers and a second checks that answer for anything unwanted. Unsuitable as soon as the steps build on each other.
Evaluator-optimizer. One agent produces, a second evaluates and gives feedback, in loops. This fits when clear evaluation criteria exist and rework measurably raises quality, for instance in translation or in code with security requirements.
The number that makes the trade-off visible
What is notable about this guide is that it presents the decision not as a ladder of progress but as a trade-off between the need for control, complexity and resources. Anthropic’s internal evaluation reports a 90.2 percent advantage for multi-agent systems over a single agent on broad tasks with several independent threads. The same systems consume roughly ten to fifteen times as many tokens.
Both numbers belong together. The added value has to justify the cost first, and whether it does is decided by the use case, not the architecture. On a task that has only one thread anyway, ten times the tokens (the billing unit for processed text) buys you nothing.
Which architecture fits law firms?
Translate these patterns into a law firm setting and the choice shifts noticeably toward the boring designs.
A deadline calculation, a conflict check, a money laundering risk check: these are processes with clear dependencies, defined intermediate states and a human sign-off at the end, and therefore exactly the non-billable processes that suit a first step. Precisely the description the sequential workflow was designed for. Trading speed for accuracy is not a concession here, it is the point of the exercise. And without traceability, admitted lawyers cannot take responsibility for the result at all.
A swarm of autonomously negotiating agents, by contrast, founders on its central price. Behaviour that is hard to predict is the very quality a case file tolerates least. There are certainly tasks in a firm where the open research variant makes sense. They just tend to sit where the result gets reviewed on the merits afterwards anyway, and not where a deadline comes out of the system.
The parallel design has a place of its own: one model produces, a second checks against criteria. That design mirrors a principle firms have always known, the four-eyes principle.
What I take away
The temptation when building is always to build the most sophisticated system you can handle. The real achievement lies in choosing the simplest one that meets the requirement, and adding complexity only once it demonstrably pays off. That too is more of a spectrum than a yes-or-no question.
The content of this blog was created partly with the support of AI.