Cortex · Reasoning Mode

An AI that works the problem through

In Reasoning Mode the model does more than fetch an answer. It plans the work, calls the tools it needs, reads what comes back, and explains what it found and what it would do about it.

01

A look-up tells you what happened. Reasoning tells you why

Most questions people ask their systems are look-ups. Where is this order? How many units are on hand? Which invoices are overdue? One tool, one answer, done.

The questions that matter in a bad week are different. On-time shipping has dropped and nobody knows why. One customer's orders keep going out short. Margin slipped in one region and held in the others. There is no single tool for questions like these, because the answer is not sitting in one table. Someone has to pull the numbers, notice what stands out, check a second system, rule a few things out, and only then say what happened.

That is the work an analyst does, and it is the work Reasoning Mode does. You ask the question the way you would ask a colleague. The model plans how to answer it, calls the tools it needs, reads what comes back, and follows the trail until it can explain what it found and what it would do about it.

02

One question, followed through

The clearest way to see the difference is to follow a single question from start to finish. This is the one at the top of the page: on-time shipping fell from 96% to 88% this month. What caused it, which customers are hit hardest, and what should we fix first?

Step 1 of 4It starts with a plan

A question like this has three parts, and none of them can be answered by a single call. So before it touches a system, the model writes down how it intends to go about it: measure the drop by site and by carrier, find what the late orders have in common, see which customers are affected, then rank the fixes by how much each one would recover.

The plan appears in the chat, so you can see how the question is being approached, not just where it ends up.

Plan

  1. Measure the drop by site and by carrier
  2. Find what the late orders have in common
  3. See which customers are affected
  4. Rank the fixes by how much they recover
Figure 1. The plan, as it appears in the chat before the first tool is called.

Step 2 of 4It measures before it guesses

The first call fetches the month's shipments: 4,212 rows. That is too many for a model to read, so the result is saved and the model asks it a question instead. What is the on-time rate at each site?

The answer narrows the problem at once. The West site is where it was last month. The whole drop is at East. Without that step the investigation would have spread across both sites and every carrier. With it, each call from here on is aimed at one place.

On-time shipping this month, by site

East81%
West95%

Last month, both sites together: 96%

Figure 2. One query on the saved shipments. The drop is at one site, not across the company.

Step 3 of 4It lets the data change its course

Now the model looks at East. The stock history shows two items sat at zero for four days. That would explain a lot of late orders, and a hurried analysis would stop there. But the plan also said to check the carriers, and the pickup records show three missed pickups at the same site in the same month.

Two causes need two fixes, so the model goes back to the saved shipments and counts the late orders by cause: 148 waited on stock, 97 on missed pickups, and 31 had other reasons. It did not know it would need that query when it started. What it read decided what it asked next. That is the part a fixed report cannot do.

276 late orders at East, by cause

1489731
  • Waited on stock148
  • Missed carrier pickups97
  • Other reasons31
Figure 3. A second query, written only after the stock and pickup records had been read.

Step 4 of 4It ends with something you can act on

The answer is short. Two causes at East explain 245 of the 276 late orders. Three customers account for 60% of them. Raise the reorder level for the two items first, because that recovers the most, then review the carrier's pickup slots.

Every figure in that answer came from a call you can open. The plan, the six tool calls and what each one returned stay on the screen above it, so anyone who doubts the conclusion can retrace it.

AnswerTwo causes at East explain 245 of the 276 late orders: two items out of stock for four days (148) and three missed carrier pickups (97). Three customers account for 60% of them.

Suggested next step: raise the reorder level for both items first, then review the carrier's pickup slots.

Figure 4. The answer. The plan and the six tool calls behind it stay on the screen above it.

03

Why the numbers can be trusted

Language models are good at reading and reasoning, and poor at arithmetic over large amounts of data. Ask one to total a million rows from memory and it will give you a confident number that may be wrong. Reasoning Mode is built so that it never has to.

When a tool returns a small result, a dozen rows for instance, the model reads all of it. When the result is large, Smart Forge saves it and gives the model a short preview with the column names. From then on the model works like an analyst with a database: it writes a query in ForgeQL to count, sum or group the saved data, and it reads the result of that query.

The model decides what to ask. The data does the arithmetic.

There is a second benefit. A large result is never pasted into the conversation, so it does not crowd out the question or run up the AI cost. And the follow-up queries are read-only and time-limited, so a question cannot change the data it is asking about.

How a result reaches the model

12 rowsThe model reads every rowAnswer
1,204,331 rowsSaved. The model gets a short previewselect site, count(*) ... group by siteAnswer
Figure 5. Small results are read. Large results are saved and queried with ForgeQL.

Read more about ForgeQL.

04

One pass, or a plan

Not every question deserves an investigation. "Where is order SO-48290?" should be answered in a moment, not planned. So the way an agent works is a setting, chosen when the agent is set up.

Answer in one pass

The agent reads the question, calls what it needs and replies. Right for direct questions with a clear answer.

Where is order SO-48290?
Plan, then work through it

The agent writes out its steps first, then follows them. Right for questions with several parts.

Why did on-time shipping fall this month?

Both stop after a set number of steps, 30 unless you change it, so a run that is going nowhere ends on its own. See how an agent is configured.

05

Freedom, inside limits

Reasoning Mode gives the model more room than a look-up does, and that is only acceptable if the room has walls. There are four.

  1. It can only use its own toolsAn agent is given a list of tools when it is set up. It cannot reach for anything outside that list, however the question is phrased.
  2. Its guardrails travel with itAn agent can be read-only. Chosen actions can wait for a person to approve them. Sensitive values, such as email addresses, can be hidden before the model reads them.
  3. It runs out of stepsEvery run has a step limit. An investigation that is not converging stops, and says so.
  4. Its work is on showEach tool call is listed with its arguments, whether it worked and how long it took. Nothing happens off the page.

Read more about guardrails.

06

When to use it, and when not to

Reasoning Mode is not the right choice for everything. Because the model reads your results, those results are sent to the model you chose. For some data that is not acceptable, and for a simple list it is more than you need.

Operational Mode is the counterpart. There the model sees only names and descriptions, never a record. It picks the tool, Smart Forge runs it, and you see the rows. It cannot investigate, because it cannot read, but nothing leaves your systems. The two are meant to be used together: Reasoning Mode for the questions that need thinking, Operational Mode for look-ups and for data that must stay where it is.

Compared onReasoning ModeThe model reads the resultsOperational ModeThe model never sees your data
What the model seesYour question and the results the tools returnNames and descriptions only. No records.
Best forWhy, what changed, what should we doLook-ups, lists, charts and routine tasks
How it answersPlans, investigates and explainsPicks the tool. Smart Forge runs it and shows you the rows.
Depth of analysisFull: it can follow a pattern across several callsLimited: it works without reading the data
Sensitive dataGuardrails can hide chosen values from the modelNo data reaches the model at all
AI costTokens for every result the model readsTokens only for choosing the tool

If you want the depth of Reasoning Mode and need the data to stay inside your network, there is a third route: run the model yourself. Bring your own model, or read about Operational Mode.

07

Where it goes from here

Reasoning Mode is also the base for other parts of Smart Forge. When a question spans several systems, a lead agent can hand parts of it to specialist agents and combine what they find. When an answer is worth keeping, it can become a report, or a dashboard tile that refreshes on its own. And the agents you configure yourself, with their own tools, skills and guardrails, work in the same way.

It can also learn how you like to work. If it notices a preference, such as how you want figures shown, it suggests remembering it. Nothing is saved until you accept.