rbergman/tiered-delegation

Tiered Delegation

Choose the right coordination mechanism for every task. Not everything needs a team, and not everything should stay in a single session.

Three Tiers

| Tier | Mechanism | When | Token cost | |------|-----------|------|------------| | Single session | Direct work | Simple tasks, <30 lines, <3 files | Lowest | | Subagents (Task()) | Fire-and-forget | Focused result-only tasks, research, lint/test, no inter-agent coordination needed | Low-Medium | | Agent Teams | Persistent teammates | Complex multi-perspective work, adversarial refinement, cross-layer coordination, debate | High |

Subagent Sweet Spots

Use Task() subagents when:

• Research and exploration — Explore agent gathers information, returns findings
• Code review that doesn't need cross-reviewer discussion
• Running quality gates — lint, typecheck, test in parallel
• Single-file implementations — contained scope, no coordination needed
• Any task where only the result matters — no need for workers to interact

Subagents are fire-and-forget. The orchestrator dispatches, waits for the result, and moves on. There is no inter-agent communication.

Agent Teams Sweet Spots

Use Agent Teams when:

• Research with competing hypotheses — investigators challenge each other's findings
• Collaborative code review — reviewers debate findings, surface deeper issues together
• Adversarial spec refinement — live debate produces better specs than sequential pipeline
• Multi-module implementation needing coordination — shared interfaces, dependency ordering
• Architecture decisions needing multiple perspectives — trade-offs require deliberation
• Any task requiring inter-agent discussion — the conversation between agents is the value

Teams are persistent. Teammates see each other's messages, react, challenge, and build on each other's work.

Planning Surface (Step 0)

Before choosing a delegation mechanism, consider whether the planning phase needs a richer review surface:

• Simple/medium tasks: Plan in the terminal as usual
• L/XL tasks with ambiguous scope: Use /ultraplan to draft the plan in the cloud. Review with inline comments and emoji reactions on claude.ai, then teleport the plan back to the terminal for /breakdown → orchestrator delegation

Ultraplan is a planning tool, not an execution tool. It doesn't replace any tier — it improves the input that feeds into delegation.

Decision Questions

Work through these in order:

1. Do workers need to talk to each other? → Teams
2. Is it result-only work? → Subagents
3. Multiple perspectives needed? → Teams
4. Sequential dependencies between workers? → Usually subagents (serial)
5. Would the work benefit from adversarial tension? → Teams

If you answer "subagents" to most but "teams" to even one, consider whether the team-worthy aspect can be isolated. Sometimes you run subagents for the bulk and a small team for the contentious part.

Cost Awareness

• Each teammate is a separate Claude instance with full context
• Teams use significantly more tokens than subagents
• Don't use teams for routine tasks — overhead exceeds benefit
• A 3-person team costs roughly 3-4x a single subagent for the same work
• Reference dm-work:orchestrator delegation thresholds as the subagent baseline

Escalation Path

Start with subagents. Escalate to teams only when needed.

1. Attempt the work with Task() subagents
2. If subagent results need discussion or synthesis beyond what the orchestrator can do alone, escalate to a team
3. If team results need broader input, expand the team or run a council

The cheapest correct answer wins. Don't reach for teams out of habit — reach for them when the problem demands collaboration.

Related Skills

• dm-work:orchestrator — Delegation thresholds and subagent launch templates
• dm-team:compositions — Reusable team templates for common workflows