Agents

Agents are configured, not developed

An agent in Smart Forge is six things you fill in: who it is, what it is told, which tools it may use, how it works, its guardrails and its skills. There is no agent code to write, test and maintain.

Six parts. All of them are settings

Each part is a step in the set-up screen. A badge shows when the agent is complete and ready to be pushed.

1

Identity

A name, an ID and a description.

The description is what gets work sent to this agent. When several agents are available, the planner reads each description to decide which one should take a step, so it says what the agent does, when to use it and when not to.

2

Instructions

The agent's standing orders, in plain language.

How it should answer, what it must always include, what it must never do. Examples can be added to show the kind of answer you expect.

3

Tools

Which tools the agent may call.

Per system, choose all tools, which includes tools added later, or select a fixed list. Tools that change data carry a "write" badge and a warning, and the screen shows how much of the model's context the selection uses.

4

Execution

How the agent goes about its work.

Answer in one pass, or plan first and then work through the plan. A limit on the number of steps stops a run that is going nowhere.

5

Guardrails

What the agent may see and do.

Read-only, approval before chosen tools run, and rules for sensitive data. Read about guardrails.

6

Skills

Written procedures the agent can follow.

A skill is a short document that explains how to do a particular job, for example how to triage a late order. The agent opens the skills that fit the request.

One agent, one definition

Everything about the agent, on one screen

When the six parts are filled in, the agent is ready to commit. Its definition is saved next to your tools and is pushed, reviewed and merged in exactly the same way, so every change to an agent has an author, a reviewer and a history.

  • No connection insideThe definition names systems, not connections. The connection is chosen when the agent runs, so the same agent works on Test and on Prod.
  • No model insideThe AI model is picked for each chat or each scheduled run, not fixed in the agent.
  • Live after the mergeOnce the change is merged, the agent can be picked in Cortex.
Skills

Teach an agent a procedure by writing it down

A skill is a short document with a name and a description. Your experts write the steps they follow themselves. The agent reads the description, and opens the full procedure only when a request calls for it.

  • Stored with the systemA skill lives in a folder next to that system's tools.
  • Shared skillsA skill placed in the Enterprise folder is offered to every agent.
  • Checked before it is pushedThe skill editor shows what the agent will see and runs checks on the file.

From set-up to on duty

Putting an agent to work has no build step and no deployment script.

  1. Fill in the six partsThe graph beside the form shows how the agent will run.
  2. Push and reviewThe agent file goes through a pull request like any change.
  3. Use it in chatAfter the merge, pick the agent in Cortex and ask it something.
  4. Put it on a scheduleIn Reflex, save a request for the agent and choose when it runs.

Developing an agent, and configuring one

DevelopingConfiguring in Smart Forge
Code for the agent loop, tool calling and error handlingThe loop is part of the platform. You choose how it runs.
Tool access wired in by a developerTick the tools the agent may use.
Safety checks written by hand for each agentRead-only, approvals and sensitive-data rules are switches.
A new release to change how it behavesEdit the instructions or a skill, and push the change.
Only developers can read what it doesA readable file and a graph that anyone can follow.