The strongest agent systems I have built do not start with "How many agents can we add?" They start with a workflow that is expensive, repetitive, slow, or difficult to audit. The architecture follows the work, not the trend.
Start with deterministic software
If a calculation can be done with code, I keep it in code. Validation, parsing, scoring rules, schema checks, data profiling, and permissions should not depend on a model unless the task genuinely requires judgment. This keeps the system faster, cheaper, and easier to debug. It also gives the agent a smaller, clearer decision surface.
Use agents for judgment and coordination
Agents become useful when the workflow needs context building, tool selection, synthesis, ambiguity handling, or multi-step reasoning. In analytics systems, that can mean selecting the chart that best explains a trend. In QA systems, it can mean reading a transcript and attaching evidence to a compliance score.
Make every agent observable
A production agent should leave a trail: input, selected tool, model, output schema, confidence, latency, token cost, and failure state. I treat traces as an engineering interface, not a debugging afterthought. They should answer what happened and make regressions comparable across prompts, models, providers, and releases.
Route models by task
Not every request deserves the largest model. A dependable platform routes simple extraction and classification tasks to efficient models, reserves stronger models for nuanced reasoning, and defines explicit retry or fallback behavior. Cost and latency are part of quality because a workflow nobody can afford or wait for is not production-ready.
Design failure paths before scale
Tool timeouts, malformed outputs, partial data, provider changes, and retrieval misses are normal system states. Each agent step needs a bounded retry policy, a structured failure result, and a clear route to deterministic fallback or human review. This is where orchestration earns its place: coordinating reliable work, not multiplying prompts.
Keep humans in the workflow
The point of enterprise AI is rarely to remove every human decision. The better goal is to move people from repetitive work into review, exception handling, coaching, and decision-making. That is why I design agent systems with evidence, redaction, approval states, and audit logs.
The production standard I use
A multi-agent system is ready only when it can answer six questions: what happened, why it happened, what it cost, how confident it was, what happens when it fails, and where a person can review the result. Agent count is not the measure of sophistication. Controlled outcomes are.