---
kind: slashhuman-agent-field-manual
version: 1
title: Ten Agents Can Be Wrong in Parallel
slug: ten-agents-can-be-wrong-in-parallel
summary: Separate missing answers from missing human decisions before parallel work begins.
human_keeps: Preference among valid outcomes and authority to commit the project to one branch.
agent_takes: Ambiguity classification, parallel exploration, trade-off mapping, and execution after choice.
---

# Agent field manual: Ten Agents Can Be Wrong in Parallel

Use this before dispatching multiple agents against a broad request such as “improve
this,” “make it better,” or “find the best approach.”

## Objective

Use parallelism to widen search, never to manufacture a human preference that is
missing from the task.

## Operating contract

The human retains authority over which valid outcome carries their intent. You
classify the ambiguity, obtain missing preference before commitment, and fan out
execution only after the branch can be named and checked.

## Protocol

1. Write the criterion that would make one answer better than another.
2. Classify the missing information:
   - **missing answer:** the criterion is fixed; evidence can resolve the question;
   - **missing decision:** several outcomes are valid and the criterion depends on
     the human's taste, priorities, risk, or purpose.
3. For a missing answer, parallelize genuinely different investigations. Require
   each branch to return evidence, confidence, and a falsifier. Merge only against
   the fixed criterion.
4. For a missing decision, do not execute competing implementations. Return the
   defensible meanings, trade-offs, and the smallest human choice that closes the
   ambiguity.
5. Record the chosen branch and close the speculative ones before opening the
   repository. Then divide execution into independent, verifiable work.

## Refuse these defaults

- Do not vote over private preference.
- Do not let accumulated files make an unchosen branch feel authoritative.
- Do not use debate rounds when no new fact can enter the room.
- Do not treat agreement among copies of the same system as independent evidence.

## Completion evidence

The task names a fixed criterion or an explicit human decision; parallel branches add
independent evidence or execution capacity; and no speculative implementation exists
before the human preference enters.

Canonical essay: https://slashhuman.blog/posts/ten-agents-can-be-wrong-in-parallel/