> ## 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.

# Lazy Discovery

> How Manowar exposes broad capability without dumping every tool, model, and agent into the prompt.

Manowar gives every agent a small default capability surface:

```text theme={null}
Use query_catalog when the tool family is unclear. Use models_call for media generation. Use connectors_find for external accounts. Use search_call for current facts. Use agent_find for agent discovery. Use swarm_propose to draft and submit a multi-agent plan.
```

The agent sees a few entry points first. It asks those entry points for the exact connector, model, agent, or search backend only when the task needs one.

## Surface Tools

| Tool              | Discovers or runs         | Backing service                               |
| ----------------- | ------------------------- | --------------------------------------------- |
| `connectors_find` | One MCP connector         | `CONNECTORS_URL /mcps/search`                 |
| `swarm_propose`   | One typed MAL plan        | In-process harness interpreter                |
| `agent_find`      | One registered agent      | `API_URL /agents/search`                      |
| `swarm_delegate`  | One registered agent call | Harness subagent engine                       |
| `models_call`     | One model call            | `MODELS_URL /search`, then internal inference |
| `search_call`     | One web search            | Perplexity, OpenAI, or Gemini backends        |

## Why It Matters

The model catalog, connector catalog, agent index, and tool schemas can be large. Dumping them into every prompt wastes tokens and makes tool selection worse. Lazy discovery keeps the prompt stable and asks the owning service for ranked candidates.

## Tool Result Budgets

Tool results are compacted before returning to the model. Defaults live in `harness/tools/index.ts`:

| Env var                          | Default |
| -------------------------------- | ------- |
| `AGENT_TOOL_RESULT_MAX_CHARS`    | `1800`  |
| `AGENT_TOOL_RESULT_ARRAY_ITEMS`  | `3`     |
| `AGENT_TOOL_RESULT_OBJECT_KEYS`  | `12`    |
| `AGENT_TOOL_RESULT_STRING_CHARS` | `240`   |
| `AGENT_TOOL_RESULT_MAX_DEPTH`    | `5`     |

Large JSON results are shortened with explicit `__compose_truncated` metadata so the model knows the result was compacted.

## Comparison

| System            | Discovery style                                                 | Manowar style                                                                  |
| ----------------- | --------------------------------------------------------------- | ------------------------------------------------------------------------------ |
| OpenAI Agents SDK | Tools are supplied by the app in code.                          | Default tools discover service-owned catalogs on demand.                       |
| OpenClaw          | Gateway and plugins expose channel and capability surfaces.     | Manowar exposes Compose services through compact runtime tools.                |
| LangGraph         | Nodes choose from state and tool bindings defined by the graph. | The native loop keeps a small tool surface and lets ranked services expand it. |

## Related

* [Tools overview](/manowar/tools/overview)
* [Model store](/manowar/tools/model-store)
* [MCP store](/manowar/tools/connectors/mcp-store)