Why this matters now
OpenAI’s July 9 release describes an ultra setting that coordinates multiple agents across parallel workstreams, while its tool-calling approach lets programs filter intermediate results before the next decision. That changes the shape of a good request: the brief must explain how work is divided and how results become one answer.
Parallel work is powerful because it reduces waiting. It is risky because it can multiply ambiguity. Four agents can produce four polished answers to four slightly different questions. The brief is the control layer.
The four rules of a parallel brief
1. Give each lane one job. “Research the market” is too broad. “Find three current competitors and cite their pricing pages” is a lane.
2. Keep evidence boundaries visible. Say which files, links, systems, or time period a lane may use. A fast lane with the wrong source is still wasted work.
3. Standardize the return. Ask every lane for the same shape: finding, evidence, uncertainty, recommendation. Synthesis becomes comparison instead of translation.
4. Reserve the final decision. The synthesizer should resolve conflicts, flag missing evidence, and state what still needs a human call. Do not let the first confident lane become the default answer.
Copy this prompt
Use this when a task is large enough to split:
Run this as four parallel workstreams, then synthesize one answer.
Shared objective:
[describe the finished result]
Lane 1 - Evidence:
[one narrow question]
Use only: [sources]
Return: finding, citations, uncertainty.
Lane 2 - Critique:
[one narrow question]
Check: [risks, gaps, contradictions]
Return: issue, evidence, severity.
Lane 3 - Options:
[one narrow question]
Compare: [criteria]
Return: option, tradeoff, recommendation.
Lane 4 - Execution:
[one narrow test or prototype]
Stop before external changes.
Return: result, failure, next step.
Synthesis rules:
- reconcile conflicts instead of averaging them
- keep citations beside claims
- label assumptions
- list unresolved decisions
- do not take external action without approval
When to use it
Use a parallel brief for research, planning, audits, code review, content packages, and any task where independent perspectives reduce blind spots. Keep a single agent for a short, tightly coupled task where every step depends on the previous one.
Start with two lanes, not four. Add lanes only when each one has a distinct question and a useful artifact to return. The point is not to create more activity. It is to shorten the path to a better decision.
