best-of-n

best-of-n generates N variants of a response in parallel using the same agent and prompt, then has a judge pick the best one.

When to use it

• Creative tasks where quality varies between runs (headlines, taglines, ideas)
• High-stakes generation where you want the best of multiple attempts
• Tasks where diversity of output improves selection
• Any situation where running the same prompt multiple times and picking the best is valuable

File structure

agents/my-best-of-n/
├── AGENT.md    ← pattern: best-of-n, n, config
└── judge.md    ← judge system prompt

AGENT.md

---
name: headline-generator
description: Generates the best possible headline for an article.
version: "1.0.0"
pattern: best-of-n
n: 5
call:
  model:
    role: writer
    temperature: 0.9
---

Generate a single compelling, specific headline for the article.
The headline should be accurate, engaging, and under 70 characters.
Return only the headline — no explanation, no quotes.

judge.md

---
call:
  model:
    role: thinker
    temperature: 0.0
---

You receive N candidate headlines for the same article.
Each is labeled with its variant number.

Evaluate each headline on:
1. **Accuracy** — does it accurately represent the article?
2. **Engagement** — is it compelling and click-worthy?
3. **Clarity** — is it specific and not vague?
4. **Length** — is it under 70 characters?

Select the single best headline.

Call finish with key="done" and ONLY the winning headline as value (no explanation).

Scaffold

tama add best-of-n my-agent

How it works

Input → N identical agents run in parallel (each with temperature for variety)
  variant-1 → headline attempt
  variant-2 → headline attempt
  variant-3 → headline attempt
  ...
          ↓
judge receives all N variants labeled by number
          ↓
judge picks the best → final output

Configuring variety

Use temperature on the AGENT.md to control how different the variants are:

• temperature: 1.0 — high variety, more creative divergence
• temperature: 0.5 — moderate variety
• temperature: 0.0 — all variants nearly identical (not useful for best-of-n)

N variants

The n field controls how many variants to generate. Typical values:

n	Use case
3	Quick selection, low cost
5	Good balance of variety and cost
10	Creative tasks, high quality bar