Platform · Architecture

A cognitive layer, not another application

Smart Forge sits between your people and the systems you already run. It connects to each system once, and every assistant, agent, screen and report works through it, with the same security, the same guardrails and the same record of what happened.

01

One layer, not more wiring

Every new assistant, report or script is usually wired to each system it needs. Each one works. Together they become a web nobody can map.

Four ways of working with four systems already means sixteen links, and the systems need links of their own. Every link has its own login, its own idea of who is allowed to do what, and its own way of failing. When a system is upgraded, each link is checked again. When someone leaves, their access is removed in each place it was granted.

Smart Forge changes the shape. Each system is connected once, to one layer. What the system can do is published there as tools, with the rules that apply to them. Chat, screens, reports and agents all work through that layer, so a fifth way of working does not need four new integrations. It needs none.

Today: point to point

Chat assistantDashboardsScriptsReports
WMSERPCRMData platform

22 integrations to build and maintain

With Smart Forge: connect once

Cortex chatForgeViewDashboardsAgents
Smart ForgeTools, agents, guardrails
WMSERPCRMData platform

4 connections, reused by everything

Figure 1. Four ways of working with four systems. Wired directly, every pair needs its own integration. Through one layer, each system is connected once.

02

Anyone can build an agent. Few can run one in production

Building an AI agent has never been easier. A capable team can have one answering questions in a week. The difficulty starts the moment it has to touch real systems, with real data, for real people.

Somebody has to connect it to each system and keep the connection alive. The agent has to act as the person who asked, not as an all-powerful service account. It needs rules about what it may read and change, a way to ask before it changes anything, and a record of everything it did. And each new capability tends to mean new code: written, compiled, built, deployed, and an application restarted.

None of that is the agent. All of it has to be built again for the next agent, and the one after. This is the work Smart Forge does once, for every system and every agent.

ConcernWith Smart ForgeIncluded, for every systemBuilding it yourselfFor every agent, for every system
ConnectionsEach system is connected once, and every tool, agent and screen reuses itA connector per system, rebuilt for each project
IdentityEvery call runs as the person who askedA shared service account, so nobody can say who did what
GuardrailsRead-only by default, approvals and sensitive-data rules, set per agentRules written into each agent
EnvironmentsDev, Test and Prod connections, chosen when the work runsSeparate builds for test and production
A new capabilityA tool is a definition and a query or script, reviewed in git and live once approvedWrite a service, build it, deploy it, restart the application
AuditOne record of every call: who asked, what ran, on which systemLogs spread across every application
The AI modelAny model, chosen for each chat or runTied to the provider the code was written for
The agent is the easy part. Everything around it is the product.

03

What the layer is made of

At the bottom are your systems, unchanged. Above them, one connection per system knows which environment it points at. The cognitive layer sits on top of the connections: the tools that say what each system can do, the domains that translate business words into fields and codes, the agents and skills that use them, and the guardrails, identity and audit that apply to every call.

At the top is wherever people work: a chat, a screen, a dashboard, an agent running on a schedule, or another AI tool connected over MCP. They all use the same tools and pass through the same rules.

Where people workEvery way in uses the same layer
Cortex chatForgeView screensDashboards and reportsDeployed agentsOther AI tools, over MCP
The cognitive layerCapabilities, rules and identity, in one place
ToolsDomainsAgents and skillsGuardrailsIdentityAudit trail

Any AI model

A provider, your cloud or your own serversSees only what each mode allows
ConnectionsOne per system
Dev, Test and ProdReused by every tool
Your systems, unchangedNothing is replaced or copied
WMSERPCRMTransportData platformYour own databases
Figure 2. The layers, top to bottom. Your systems stay as they are; the layer above them is where capabilities, rules and identity live, and it is the only part that talks to the AI model.

The AI model sits outside the layer and can be swapped. It can be a hosted model from any provider or one that runs in your own cloud or on your own servers, chosen for each chat or run. Read about bringing your own model.

04

Adding a system is a set of tools, not a project

Common systems are supported out of the box. When you run something we have not seen before, we build the integration with you. Either way, what is added is small: the connection and the tools that describe what the system can do.

Everything else is already there. The new system gets the same identity rules, the same guardrails, the same environments and the same audit as every other system on the layer. Nobody builds those again for it. A tool is a definition and a query or a script. It is reviewed and approved in git like any change, and once approved it is available to the whole organisation.

Adding a system

  1. Connect itCommon systems are supported out of the box. For the rest, we build the integration with you.
  2. Publish its toolsWhat the system can do is written as tools, then reviewed and approved in git.
  3. Use it everywhereChat, agents, screens, reports and other AI tools can use the new tools at once.

Already handled by the platform: connection, identity, guardrails, environments and audit.

Figure 3. A new system is a new set of tools on the same layer, not a new project.

05

From answers to decisions

The questions a business asks range from a single fact to a decision. A shared layer can serve all of them, because each kind of question uses the same tools in a different way.

The first kinds are answered directly. The model picks a tool, the tool runs, and the answer is exact and repeatable. Further up, the questions need judgement: the model has to plan, read results from several systems, and explain what they mean. That is what Reasoning Mode and multi-agent orchestration are for.

  1. Transactional“What is the status of order 48290?”
    01
  2. Contextual“Why is this order on hold?”
    02
  3. Predictive“Which orders will miss their date this week?”
    03
  4. Economic“What revenue is at risk, and where?”
    04
  5. Causal“Which decision caused this delay?”
    05
  6. Prescriptive“What should we do, and in what order?”
    06
Operational Mode: instant and repeatableReasoning Mode and multi-agent orchestration: judgement across systems
Figure 4. Six kinds of question, from a single fact to a decision. The first two are answered directly; the rest need the model to reason across systems.

06

Where your data goes, and where it does not

The layer runs inside your environment. Your systems and your connections stay there. What reaches the AI model depends on the mode, and in both modes it is decided by the layer, not by the model.

In Operational Mode the model receives the question and descriptions of the tools, never data. It answers with what to run, and that runs inside your environment, as you, with the results going straight to your screen. Because what to run is saved, the same request can run again later with no AI at all.

You

1Ask

6Result

Table or chartpin to a dashboard

Inside your environment

Smart Forge

Tool descriptionsDomainsRunnerYour identityDashboardsAudit
4Runs as you↓↑Results5
WMSERPData

Your data never leaves. Running it again needs no AI at all.

  1. 2Question and tool descriptions only
  2. 3What to run

Model provider

Any AI modelwrites what to run, nothing else

Sees: the question and the tool descriptions. Never your data.

Figure 5. Operational Mode, step by step. The model decides what to run; everything that touches your data happens inside your environment.

In Reasoning Mode the model plans the work, so it chooses the tools and their arguments. The tools run inside your environment, as you, on the chosen connection. The model then sees the results, because it has to reason about them, but only after the guardrails have masked the sensitive values. It drafts the answer, and you receive it as text, a chart or a PDF report.

You

1Ask

8Result

Answer or chartor a PDF report

Inside your environment

Smart Forge

Agent instructionsDomainsAllowed toolsGuardrailsTablesYour identity
4Runs as you↓↑Results5
WMSERPData

Connections never leave.

  1. 2Question, instructions and allowed tools
  2. 3Tools to call, with arguments
  3. 6Results, after the guardrails
  4. 7The answer, drafted

Model provider

Any AI modelplans and reasons

Sees: your question, the tools it called and their results, after the guardrails.

Figure 6. Reasoning Mode, step by step. The model sees results so it can reason about them, but only after sensitive values are masked.

07

Work that runs on its own

The same layer runs agents that nobody has to ask. A deployed agent is given a standing goal and a schedule. It wakes up, reads what it needs, decides what matters, and acts through the same tools and guardrails as a person in a chat.

1Wake up and re-checkThe agent starts itself, on its schedule, day or night.
2Detect and quantifyIt reads the warehouse and the ERP, then sizes the risk.
Standing goal“Keep late orders and lost revenue as low as possible.”
3Recommend and prioritiseRanked actions, with the reason for each.
4Act and recordCalls existing tools within its guardrails, and records what it did.

then the loop runs again

Figure 7. A deployed agent works toward a standing goal, then the loop runs again.

08

Five gates between a request and an action

A request in plain language can mean several things. Before anything runs, the layer makes sure the action is allowed, understood, attributed to a person, aimed at the right environment, and recorded.

Request

“Release the urgent orders”

Gate 1

ToolsIs this allowed?Only the agent's own tools. Read-only by default.

Gate 2

DomainsWhat did they mean?Business words map to fields, codes and arguments.

Gate 3

IdentityWho is acting?Runs as the person who asked.

Gate 4

EnvironmentWhere does it run?On the chosen Dev, Test or Prod connection.

Gate 5

SafetyWhat happened?Approvals, sensitive-data rules and a full audit trail.
Safe, recorded action

Chat, deployed agents and other AI tools over MCP all pass the same five gates.

Figure 8. Natural language is ambiguous. The gates make sure every action is allowed, understood, attributed, aimed at the right environment and recorded.

09

Inside your own applications, too

The layer is not only for Smart Forge's own chat and screens. Your own applications and portals can use its tools over MCP, the open standard AI tools use to talk to each other. A view in your application calls a tool and shows live data. A button calls a tool that changes data. Alongside your application, deployed agents keep working in the background.

Nothing is rebuilt for this. The tools and connections your application uses are the same ones everything else uses.

Your application

Smart Forge

Your systems

Orders at riskERP tool
ERP tools
ERP
Stock on holdWMS tool
Release holdWMS write tool
WMS tools
WMS
Late shipmentsTMS tool
TMS tools
TMS

Each tool runs on its system through that system's own connection.

  • Live, not batchEach view calls a tool when it is opened. One token opens one system's tools.
  • Act from the viewTools that change data, such as release or allocate, run with your application's own connection token, and every call is logged.
  • Work in the backgroundAlongside your application, deployed agents keep watching, follow up and report.
Figure 9. Your own applications use the same layer over MCP: one token per system, the same tools, on the same connections.