Skip to content
AI Learning Playground
LinkedInX
English
English

Delegation Patterns

About 5 minutes

Shiori
Target audience: Those deciding what work to split out to a subagent, those looking for delegation patterns
Prerequisites: Basic understanding of What Are Subagents

Delegation comes in a few shapes. Naming the shape makes it easier to decide what to pass in and what to expect back.

Three Basic Patterns

PatternPurposeFormat returned
ResearchCover a wide area and reach a conclusionSummary of conclusion and evidence
ReviewCheck a deliverable from another viewpointA list of findings
ParallelRepeat the same processing across targetsPer-target results
This table scrolls horizontally. Keyboard users can focus the table and use the left and right arrow keys.

Research Delegation

This pattern pays off the most. Research reads a lot, and keeping all of it in the main context crowds out later work.

graph LR
    Main["Main"] -->|scope and criteria| Sub["Subagent"]
    Sub -->|conclusion and evidence only| Main
  • Pass in: the purpose, the scope, and the criteria for judgment
  • Receive: the conclusion, the supporting locations, and anything left undecided

A format of “report everything seen” defeats the purpose. Ask for the conclusion first, with evidence kept tight.

Review Delegation

An author reviewing their own deliverable reproduces the same blind spots. Give a separate subagent the deliverable and the criteria, and nothing else.

  • Pass in: the deliverable, the evaluation criteria, the scope
  • Do not pass in: the history and intent behind its creation, which becomes bias

Withholding the creation history is what preserves the independent viewpoint.

Parallel Delegation

This applies the same processing across several targets. It requires the targets to be independent, so it does not fit work where each step depends on the previous result.

  • Pass in: the shared procedure, the assigned target, the output format
  • Watch out: without a uniform output format, the results cannot be merged afterward

Four Elements Every Instruction Needs

Whatever the pattern, an instruction should carry four things.

  1. Purpose — why this work is being done
  2. Input — the target and what information may be consulted
  3. Completion condition — where the work ends
  4. Output format — the shape of the result

The last two are the ones most often missing. Without them, the main agent cannot use what comes back.

Summary

  • The three basic patterns are research, review, and parallel
  • Research delegation should return the conclusion and evidence, not everything seen
  • Review delegation withholds creation history to keep the viewpoint independent
  • Every instruction needs purpose, input, completion condition, and output format
Quiz