> ## Documentation Index
> Fetch the complete documentation index at: https://docs.compose.market/llms.txt
> Use this file to discover all available pages before exploring further.

# Harness

> Typed multi-step execution with the Manowar Agent Loop (MAL).

The Manowar harness runs Manowar Agent Loop (MAL) plans. A MAL plan is YAML or JSON with a closed set of operations. The interpreter validates the plan, executes each step, stores saved values, resolves references, and returns a structured run result.

Use the harness when a task needs multiple steps, parallel branches, registered subagents, direct tool calls, checkpoints, proof bundles, scratch state, or shared swarm state.

## Why MAL Exists

Prompt-only agent loops spend tokens re-explaining how to act. MAL makes execution a typed runtime concern:

| Concern       | Prompt-only loop           | MAL harness                                                       |
| ------------- | -------------------------- | ----------------------------------------------------------------- |
| Step dispatch | Model decides from prose.  | Interpreter dispatches a typed `op`.                              |
| References    | Model copies text around.  | Runtime resolves `{{stepId.path}}`.                               |
| Budgets       | Usually prompt-instructed. | Runtime enforces token, tool batch, wall time, and depth budgets. |
| Parallelism   | Model describes it.        | Runtime runs `fanout` branches with concurrency controls.         |
| Resume        | App-specific.              | ValKey checkpoint store can resume by `runId`.                    |

## Minimal Plan

```yaml theme={null}
description: "Summarize a source and stop"
steps:
  - op: tool
    saveAs: source
    toolName: search_call
    args:
      query: "Compose.Market Manowar memory"

  - op: synthesize
    saveAs: summary
    from: ["source"]
    model: "gpt-5.5"
    instruction: "Summarize the result in five bullets."

  - op: stop
    output: "{{summary}}"
```

## Ops

| Op           | Runtime behavior                                |
| ------------ | ----------------------------------------------- |
| `task`       | Spawn a focused registered subagent.            |
| `delegate`   | Call a registered deployed agent.               |
| `fanout`     | Run branches in parallel.                       |
| `tool`       | Call a tool directly.                           |
| `if`         | Branch on a safe condition.                     |
| `loop`       | Repeat a bounded block.                         |
| `scratch`    | Read/write private run state.                   |
| `synthesize` | Combine saved artifacts with an explicit model. |
| `stop`       | Return a final output.                          |
| `ask_user`   | Ask the host for input.                         |

## Comparison

| System     | Multi-step control                            | Manowar harness                                                   |
| ---------- | --------------------------------------------- | ----------------------------------------------------------------- |
| LangGraph  | State graph with nodes and durable execution. | Typed plan interpreted inside the Compose runtime.                |
| Google ADK | Workflow agents and runner services.          | Registered Compose agents plus MAL ops.                           |
| AutoGen    | Event-driven agent teams and messages.        | Runtime-level plan dispatch, budgets, and proof/checkpoint hooks. |

## Related

* [Harness architecture](/manowar/harness/architecture)
* [Planning](/manowar/harness/planning)
* [Checkpoints](/manowar/harness/checkpoints)