What is AI agent orchestration?

Horme is being built around a simple idea: using more AI agents is only useful if the work between them stays connected.

A developer might ask Claude Code to implement a feature, Codex to review it, Claude Code to fix the findings, and Codex to verify the result. Without an orchestration layer, the developer becomes the routing system. They copy context, watch terminals, remember what depends on what, move review findings between tools, and decide when the next step can start.

Horme is designed to move that coordination into the product.

The problem is not access to more models

Developers already have access to strong coding agents. The coordination problem appears when several of them take part in the same job.

claude code   implements auth
              you copy the context
codex         reviews auth
              you copy the findings
claude code   fixes the findings
              you check the status
codex         verifies

The developer is doing useful judgment work, but also a large amount of mechanical routing. With orchestration, the intended flow becomes one workflow:

The core objects

These objects let Horme represent work as a connected system instead of a collection of unrelated chats or terminals.

Each of these is described in more detail in How Horme works.

One job can branch

Agent work is not always sequential. A feature can be implemented once, then reviewed and tested in parallel before the findings converge.

Human intervention is part of the system

Orchestration does not mean removing the developer from the loop. Some decisions belong to the human.

agent needs permission
        ↓
workflow marked as needing you
        ↓
you review the request
        ↓
approve or reject
        ↓
only affected downstream work waits

Everything unrelated should keep moving. The principle is simple:

Downstream work never moves past an unresolved intervention. Everything unrelated keeps moving.

In Horme this becomes the Needs You inbox, which is being built.

Why use different agents?

Horme does not assume one model or one coding harness will be best for every step.

Different systems can have different strengths in implementation, review, debugging, speed, cost, tool use, context handling, or developer preference. Even when one system is the default, an independent second agent can be useful for review or verification.

The value of orchestration is not the number of logos on the screen. It is the ability to connect the work. Our longer argument is in Why we think AI development will become multi agent.

What Horme is building toward

Horme starts with Claude Code and Codex.

The long term goal is a control layer where developers can define work once, assign the right agent to each step, let results move between steps automatically, intervene when judgment is required, and inspect the full history afterward.

The interface should answer six questions quickly:

• What is working?
• What is waiting?
• What needs me?
• What finished?
• What did each agent produce?
• What happens next?

That is AI agent orchestration. The patterns it enables are covered in Multi agent coding workflows.