Multi agent coding workflows

The simplest version is not several agents running at the same time. It is several agents taking part in one coherent piece of work.

The patterns below are the workflows Horme is being built for. Today you can run each step as its own task on the agent you choose; the automatic connections between steps are in development.

Pattern 1: Implement, review, fix, verify

This is the core Horme workflow.

Pattern 2: Fan out and fan in

Some work benefits from parallel checks. Three checks can begin once the implementation is available, and the fix begins only when all of them are complete.

Pattern 3: Independent verification

Sometimes the value of a second agent is not specialization. It is independence.

Horme should preserve enough context for the reviewer to understand the work without forcing it to adopt the previous agent's conclusions.

Dependencies

Dependencies determine what can start.

task B depends on task A

A working     B waiting
A completed   B released

A workflow can therefore express both safety and concurrency.

A must finish before B
A does not affect C

if A waits for a human:
  B waits
  C keeps moving

That is more useful than pausing an entire project because one agent needs attention.

Handoffs

A handoff is the connection between two pieces of work. A good handoff should answer:

• What was the goal?
• What changed?
• What output matters?
• What did the previous agent discover?
• What does the next agent need to do?
• What should not be repeated?

Horme is designed to attach this context to the work itself instead of relying on the developer to copy it manually. The objects involved are described in How Horme works.

Artifacts

Artifacts are durable outputs, such as:

• A code diff or commit
• A review report
• A test or verification result
• A generated file, screenshot, or patch

Agent workflows need something concrete to pass forward. A handoff should point to the artifact that matters, not merely say that another agent finished.

Interventions

The developer should be interrupted when judgment is required, not whenever an agent changes state. For example:

• Approve a permission
• Choose between two implementations
• Answer a missing requirement
• Review a risky change
• Approve the final workflow

The workflow should preserve that decision as part of its history.

Provenance across agents

A multi agent workflow becomes difficult to trust if the final result cannot be traced. Horme keeps the chain connected:

task → run → agent → artifact
  → handoff → review → next run
  → human approval

The principle

The goal is not to maximize the number of agents. The goal is to make work move through the right agents without making the developer become the coordination layer.

One job. Several agents. Context stays connected.

For why this matters beyond a single workflow, read What is AI agent orchestration?