task-delegation

Files (1)

• SKILL.md

SKILL.md (first lines)

---

## Task Delegation

Delegate independent sub-tasks to background child agents using the `task` tool, then collect results with `wait`.

### The `task` Tool

**Parameters:**
```
task(description="label", prompt="instructions", agent_type="explorer", model="optional", reasoning_effort="optional")
```

| Parameter | Required | Description |
|-----------|----------|-------------|
| `description` | Yes | Short label shown in UI and logs |
| `prompt` | Yes | Full instructions for the child agent |
| `agent_type` | Yes | Type hint; defaults to "explorer" if empty |
| `model` | No | Override model for this task |
| `reasoning_effort` | No | "low", "medium", or "high" |

**Returns:** A result with `agentID` in metadata:
```json
{
  "Output": "delegated X to Y",
  "Metadata": {
    "agentID": "task:parent-session:uuid",
    "agentName": "Curie",
    ...
  }
}
```

### The `wait` Tool

**Parameters:**
```
wait(agent_ids=["task:parent-session:uuid"])  // optional
```

If `agent_ids` is omitted, waits for **all** outstanding child agents.

**Returns:**
```json
{
  "Output": "finished waiting for N agent(s)",
  "Metadata": {
    "count": 2,
    "results": [
      {
        "id": "task:...",
        "name": "Curie",
        "status": "completed",
        "summary": "child agent output..."
      }
    ]
  }
}
```

### Identifying Parallelizable Sub-tasks

Sub-tasks are parallelizable when:
1. They do not depend on each other's output
2. They can start with information available now
3. Results can be merged afterwards

Good candidates:
- "fetch these 5 URLs independently"
- "run these 3 searches in parallel"
- "analyze each of these 4 log files"

Bad candidates (sequential required):
- "debug X, then write a fix" — fix depends on debug result
- "check auth, then access data" — second depends on first

### Writing Clear Task Prompts

The child agent receives **only** what is in `prompt`. Write complete, self-contained instructions:

**Bad:**
```
check the logs
```

**Good:**
```
Read /var/log/app.log. Find all ERROR-level lines from the past 24 hours.
Return a JSON array with fields: timestamp, message, stack_trace (if present).
```

Include:
- Exact file paths or URLs
- Expected output format
- What to return (the `summary` field in wait results)

### Child Agent Capabilities

Child agents have **limited tools**:
- Disabled: `message`, `spawn`, `task`
- Enabled: file operations, grep, read, web_search, web_fetch, etc.

Children **cannot send messages** to users or spawn further agents.

### Iteration Limit

Children are limited to **15 iterations** max. Keep tasks focused enough to complete within this limit.

### Handling Results

Check the `count` field in wait results. Iterate over `results`:

```go
for _, r := range results {
    switch r.Status {
    case "completed":
        // use r.Summary
    case "error":
        // handle r.Error
    case "running":

From Nomadcxx/smolbot, BSD-3-Clause.

More from Nomadcxx/smolbot