Skill: Delegation
"A senior engineer knows when to code, when to delegate, and when to collaborate."
When
- At blueprint creation: Assess each task and assign Owner (
self, sub-agent, or team).
- Before each task: Re-assess. Context may have changed.
Mechanism First
The epigraph says "when to code" — but that is not a delegation choice. ground-mechanism decides
the executor kind (code or model). This skill decides the executor context (self, sub-agent,
team).
Ask mechanism first. If the answer is code, there is nobody to delegate to.
The Primitives
| Primitive |
Communication |
Best For |
| Self |
— |
Judgment-heavy, context-dependent, architectural |
| Sub-agent |
Reports back only |
Bounded, mechanical, verifiable |
| Agent team |
Peer-to-peer |
Research, competing hypotheses, cross-layer work |
The Criteria
Sub-agent — all four must be true:
| Criterion |
Question |
| Bounded |
Can you specify input and expected output in one sentence? |
| Context-free |
Does it NOT require knowledge of other in-flight tasks? |
| Mechanical |
Is the approach obvious once started? |
| Verifiable |
Can success be checked without judgment? |
All yes → sub-agent. Any no → check if team fits.
Agent team — when the work needs dialogue:
| Signal |
Example |
| Competing hypotheses |
Investigate the root cause in parallel |
| Cross-layer coordination |
Frontend, backend, and tests move together |
| Research with challenge |
Findings should be questioned, not just collected |
| Peer review |
Multiple lenses on the same artifact |
If none → self.
Deeper
|
|
| authority |
privilege attenuation, and the authority gradient |
| briefings |
what a sub-agent is told, and what a team is told |
| running |
the workflow, and running in parallel |
| examples |
delegations that worked, and one that did not |
The Anti-Patterns
- Over-delegating: Spawning agents for trivial single-file changes.
- Under-delegating: Drowning in mechanical work instead of leading.
- Vague briefs: No clear success criteria.
- Fire and forget: Not reviewing output before committing.
- Scope inflation: Broader access than the task requires.
- Blind obedience: Executing flawed briefs without questioning. The
Challenge line prevents this.
- File collisions: Two teammates, same file. Split the work.
1---2name: ground-delegation3description: Strategic delegation. When to lead, delegate, or collaborate.4---56# Skill: Delegation78> "A senior engineer knows when to code, when to delegate, and when to collaborate."910## When11121. **At blueprint creation**: Assess each task and assign Owner (`self`, `sub-agent`, or `team`).132. **Before each task**: Re-assess. Context may have changed.1415## Mechanism First1617The epigraph says "when to code" — but that is not a delegation choice. `ground-mechanism` decides18the executor **kind** (code or model). This skill decides the executor **context** (self, sub-agent,19team).2021Ask mechanism first. If the answer is code, there is nobody to delegate to.2223## The Primitives2425| Primitive | Communication | Best For |26|----------------|------------------|---------------------------------------------------|27| **Self** | — | Judgment-heavy, context-dependent, architectural |28| **Sub-agent** | Reports back only | Bounded, mechanical, verifiable |29| **Agent team** | Peer-to-peer | Research, competing hypotheses, cross-layer work |3031## The Criteria3233### Sub-agent — all four must be true:3435| Criterion | Question |36|------------------|------------------------------------------------------------|37| **Bounded** | Can you specify input and expected output in one sentence? |38| **Context-free** | Does it NOT require knowledge of other in-flight tasks? |39| **Mechanical** | Is the approach obvious once started? |40| **Verifiable** | Can success be checked without judgment? |4142All yes → `sub-agent`. Any no → check if `team` fits.4344### Agent team — when the work needs dialogue:4546| Signal | Example |47|------------------------------|---------------------------------------------------|48| **Competing hypotheses** | Investigate the root cause in parallel |49| **Cross-layer coordination** | Frontend, backend, and tests move together |50| **Research with challenge** | Findings should be questioned, not just collected |51| **Peer review** | Multiple lenses on the same artifact |5253If none → `self`.5455## Deeper5657| | |58|---|---|59| [authority](authority.md) | privilege attenuation, and the authority gradient |60| [briefings](briefings.md) | what a sub-agent is told, and what a team is told |61| [running](running.md) | the workflow, and running in parallel |62| [examples](examples.md) | delegations that worked, and one that did not |6364## The Anti-Patterns6566- **Over-delegating**: Spawning agents for trivial single-file changes.67- **Under-delegating**: Drowning in mechanical work instead of leading.68- **Vague briefs**: No clear success criteria.69- **Fire and forget**: Not reviewing output before committing.70- **Scope inflation**: Broader access than the task requires.71- **Blind obedience**: Executing flawed briefs without questioning. The `Challenge` line prevents this.72- **File collisions**: Two teammates, same file. Split the work.
attac-t