Summary

Agentic leverage comes from a complete delegation loop: onboard the agent, hand off bounded work, review the output, then turn repeated handoffs into reusable process.

This is different from prompting. Prompting asks for an output. Delegation creates a repeatable operating relationship between operator, context, task, and review.

Key points

• Delegation starts with context. An agent needs the same things a new teammate would need: goal, audience, voice, source of truth, constraints, examples, and definition of good.
• The handoff must be bounded. The operator should specify the task, decision owner, expected artifact, review standard, and stop condition.
• Review is part of the work. Delegating without review is abdication, especially for claims, recommendations, external actions, and identity-linked work.
• Under-delegation and over-trusting are opposite failure modes. One keeps work stuck on the operator; the other ships unverified output.
• Repeated delegation should become infrastructure: a skill, runbook, project, checklist, context pack, or agent-facing CLI.
• Parallel orchestration is earned. Multiple concurrent agent workstreams are useful only after handoffs and review gates are reliable.

Operational relevance

Use this pattern when a workflow is repeated often enough that ad hoc prompting becomes drag.

A good delegation loop has five parts:

1. Onboard: provide durable context and decision rules.
2. Delegate: send a bounded task with a clear output contract.
3. Review: check facts, judgment, voice, risks, and completion.
4. Capture: save reusable context, examples, and corrections.
5. Orchestrate: run parallel work only when the loop is stable.

This pattern complements:

• Artifact-First Handoffs, because the handoff and review need durable artifacts.
• Planner Executor Separation, because delegation works best when planning context and execution context are intentionally scoped.
• Context Manifests and Governed Retrieval, because onboarding should be explicit rather than dependent on chat memory.

Failure modes

• Prompting instead of delegating: the operator keeps revising output manually and never creates a reusable process.
• Identity lock: the operator refuses to delegate writing, strategy, or judgment because those tasks feel like proof of personal value.
• Blanket trust: the operator stops applying critical thinking because the model usually sounds right.
• Premature parallelism: many agents run at once, but the system lacks artifacts, stop conditions, and review gates.

Open questions

• What is the minimum viable delegation card for local OpenClaw subagents?
• Which recurring workflows deserve skills versus simple reusable prompts?
• How should review gates vary by risk level?

Sources

• wiki/sources/the-art-of-ai-delegation.md