An atlas of agentic patterns.
52 ways to build with LLMs, from one augmented call to teams of agents. Each pattern runs live: watch prompts, tool calls and loops move through it, then open it for the build prompt and prompt templates.
All steps
Fails twice, then passes
Passes first time
Parts of this pattern
- Task Input / output
- The request that starts the loop, such as 'Write a board memo on the Q3 churn spike'.
- Generator LLM call
- Writes the first draft from the task, then rewrites it using the evaluator's critique each round.
- Evaluator Evaluator
- Judges each draft against the criteria and returns JSON: passed, plus a critique when it fails.
- Output Input / output
- The accepted draft, returned only once the evaluator says it meets every criterion.
- Task → Generator
- The task goes to the generator; the criteria are kept for the evaluator.
Task: Write a one-page board memo on Q3 churn. Criteria: cite a source for every number; formal tone.
- Generator → Evaluator
- Each draft is sent to the evaluator, which judges it against the criteria.
1. Churn hit 4.1% in Q3 (Finance dashboard). 2. Folks mostly bailed over the new pricing. 3. We expect recovery by Q1 (CS forecast).
- Evaluator → Generatorrevise
- The evaluator's critique goes back so the generator can fix exactly what failed.
{"passed": false, "critique": "Claim 2 has no citation; tone too casual for a board memo."}
- Evaluator → Outputpass
- When passed is true, the draft leaves the loop unchanged.
{"passed": true, "critique": ""}
Evaluator-Optimizer4 nodes · 6 LLM calls · loop · branch · 2 scenariosOpen this pattern
52 patterns · 9 families
The atlas
Every pattern sits on an island with its family. Far out, the map reads as clusters; close up, each diagram runs live.Scroll over the map to zoom, drag to move, click a pattern to open it. With the map focused, the arrow keys walk from pattern to pattern and Enter opens one.Pinch to zoom, drag to move, tap a pattern to open it.