Skip to main content
Building on Forge means deciding three things: who does the work (an agent, or a person in a position), what that worker can reach (skills, tools, procedures, knowledge and outside services), and how the work runs (a workflow graph you publish). Each guide below covers one of those decisions. Read Key concepts first if the terms are new, or start from the Quickstart to see all three end to end.

Define the worker

Define an agent

The definition’s layers, starting from a template or blank, and what makes it ready to run.

Talk to agents

Put chat in your product: channels, sessions and personal agents.

Give it skills, tools and knowledge

Author a skill

Package reusable instructions an agent loads always or on demand.

Author a procedure

Write governed code with typed inputs and outputs that runs in a sandbox.

Built-in procedures

The procedures Forge ships, with their inputs, outputs and errors.

Tool sets

The code-shipped tool sets an agent can be granted.

Knowledge in agents

What an agent can retrieve and cite, and where memory stops.

Synchronized sources

Sync a wiki or document store into knowledge in one scope.

Compose the workflow

Compose a workflow

Node types, typed inputs and outputs, and {{workflow.x}} references.

Use the Architect

Describe a workflow in words and review the graph it drafts.

Synthesis

How a missing skill, agent or procedure is created for review.

Author a judge

Add a rubric-backed judge that grades work and returns a verdict.

Connect outside services

Connect an external service

Reach a vendor such as Linear or GitHub through a provider and its credential.

Bind node credentials

Choose whose credential each connection a workflow node opens uses.

Configure a sandbox

How sandbox VMs are provisioned and how command policy applies.
When a workflow is published, Operate covers how its runs start, are governed and are observed.