Add a step. Change nothing else.

Most approaches to multi-agent automation require you to define the workflow before you run it. You map out the steps, connect the nodes, and deploy the coordination logic. When you need to change something, you change the workflow definition and redeploy.

This works when the process is fixed. It becomes a constraint when it is not.

Forge works differently. Agents coordinate through content state, not through each other. One agent produces a result and publishes it. The next agent picks it up when it appears. No direct connection between them. No workflow definition to maintain.

What that looks like

An agent receives a brief and writes a technical specification. When the specification is published, a second agent picks it up and breaks it into tasks. When each task is published, a third agent implements it.

Three agents. No handoff protocol between them. No shared memory. No orchestrator watching over the process. Each agent knows only what triggers it and what it produces. Forge handles the rest.

At any point in this process, a person can step in. Review a specification before the task-planner sees it. Edit a task before implementation begins. Archive something that went in the wrong direction. Whether you want a fully automated pipeline or a reviewed one at every step is your decision. The architecture does not prescribe it either way.

Adding a step

If you want to add a review stage between specification and implementation, you publish a new agent job that subscribes to specifications and produces reviewed specifications. The other agents do not change. The coordination layer does not change, because there is no coordination layer.

Removing a step works the same way. Archive the agent job. It stops. The rest of the workflow continues unchanged.

Full visibility at every stage

Every step in a multi-agent workflow produces content in Forge. That means every step has a complete record: what was produced, when, by which agent, from what previous state.

There is no black box in the middle managing handoffs. The workflow is the set of agent jobs and the content they produce. You can see all of it at any time.

Each agent has exactly the access it needs

Every agent job operates with its own credentials and its own access level. An agent that writes specifications can create and publish specifications. It cannot touch tasks or implementation results. An agent that implements tasks can only operate on tasks.

The boundaries are defined once and enforced by the framework. They are not managed by the agents themselves or negotiated at runtime.

*See Your AI is part of the team for how individual agent jobs work.* *See forge-agent for wiring and configuration.*