LLMs answer. Agents decide.

There was a collective fascination with large language models, and it was justified. For the first time a machine held a conversation in natural language, wrote, summarised and explained complex concepts with convincing fluency. Companies rushed to adopt. And then came the quiet disappointment, the kind nobody puts in the success case. The brilliant model knew nothing about the company using it.

That is the underlying limitation, and it is not solved by a bigger model. A market LLM knows the entire world and nothing about your operation. It can discuss inventory management in theory, but it does not know how much you have in stock right now. It can explain the concept of cash flow, but it does not know what your cash flow is at this moment. It works on broad, frozen knowledge, captured at the moment it was trained, disconnected from the living reality of the business asking the question. For conversation, it is extraordinary. For deciding about your company, it is blind.

The difference between answering and deciding is the difference between knowing and acting on what you know. Answering is returning well-articulated general knowledge. Deciding is reasoning over real, specific, current data and reaching a conclusion that becomes action. An LLM answers with what it learned about the world. An agent connected to the operation decides with what is happening in the company now. They are different categories of tool, and confusing the two is the origin of much corporate frustration with AI.

The metaphor Mars uses for this is synapses. An agent's intelligence is not formed on the generic internet. It is woven over the client's data, the ERP, the CRM, supply, finance. The generic internet offers shallow, out-of-context knowledge about any subject. The company's ecosystem offers real synapses, connections formed from the actual operation. An AI trained on the world knows a lot about everything and nothing about you. An AI connected to your operation knows what matters, which is the real state of your business.

There is a second distinction, as important as the first. An LLM operates in a single pass. You ask, it answers with what it has, and it stops. An agent operates in a loop. It receives the question, investigates, calls specialised tools, queries the data, realises it needs one more cross-reference, searches again and repeats the process until the investigation is complete. It is not a ready-made answer spat out at once. It is reasoning that develops, the way a good analyst would if they had instant access to everything and infinite patience. And the result is not only text. It is analysis and charts generated on demand, in the format the question calls for.

An example makes the difference tangible. Imagine a CFO who wants to understand why the margin on a product line fell last quarter. From the generic LLM he receives a competent lecture on the factors that usually pressure margin, input cost, sales mix, exchange rates, all correct and none useful, because it does not speak about his company. From the connected agent he receives something else. The agent goes to the real data, cross-references input cost with sales volume by region, identifies that the drop is concentrated in one specific channel because of a discount policy nobody reviewed, and shows it in a chart generated at that instant. The first explains the world. The second solves the problem.

Honesty about what each thing is good at is required. Large language models are fundamental pieces, and Mars uses Claude technology, from Anthropic, as part of its own reasoning engine, as a Preferred Services Partner in Anthropic's partner network. The question was never the model versus the agent. The model is the linguistic brain. The agent is the system that gives that brain governed access to the company's living data, memory of the business context and the ability to iterate until it decides. Without the model, the agent does not reason. Without the agent, the model knows nothing about you. Corporate value comes from the union of the two, with governance in between.

For anyone deciding on technology at a serious company, the question to ask is not which model is the smartest of the season. That race changes leader every few months and matters less than it seems. The question is another one. Is your AI connected to the living data of your operation, or is it just holding a good conversation about the world out there. One answers. The other decides.

Mars built Signals to be the second. Three specialised agents, in data, finance and supply chain, one unified brain, all thinking with your company's data. Other AIs show what they know about the world. Signals thinks with your company's data. If the distance between what your AI knows and what your company needs to decide is still wide, it is worth a conversation about closing it.