About

How to read the sky

Node shapes

• ioInput or output: where a request enters or a result leaves.
• llmA single LLM call with a specific prompt.
• toolA tool or API the model can call.
• routerA classifier or code branch that picks a path.
• evaluatorA judge: tests, a validator or an LLM grader.
• aggregatorCombines several results: a vote, a merge or a judge.
• memoryA store: vector DB, plan, scratchpad or long-term memory.
• humanA person in the loop who approves, edits or rejects.
• guardA guardrail check that can block the flow.
• agentA full agent: an LLM in a loop with its own tools.

Edges and tokens

• Solid line: a fixed path that runs every time.
• Dashed line: a return path or optional route.
• Dashed arc with a red label: a loop back to an earlier step.
• Comet tokens: messages in flight, with a short label for what they carry. The edge they travel turns red while it is active.
• Hover or focus any node or edge for what it does; edges also show an example payload.
• Sound is off unless you turn it on with the speaker button in the header. Then the diagram you are looking at plays a soft tone for each kind of token, a chord when a run finishes and a low thunk when something is rejected. Only that one diagram makes sound, and nothing plays while the tab is hidden. The same button stops all sound at once.

Methodology

Patterns are curated from Anthropic's Building effective agents, OpenAI's A practical guide to building agents, and the original papers cited on each page. Diagrams play scripted traces that illustrate control flow; they are not live model runs. Each build prompt is provider- and language-neutral on purpose: paste it into your own AI coding assistant and it uses the stack you already have.

Made by

Built by Joran Kikke. GitHub · Source

Credits

Built with Astro, ELK, Expressive Code and Pagefind. Typeset in Archivo and JetBrains Mono.

Licence

Code is MIT. The pattern write-ups, diagrams and prompts are CC BY 4.0: reuse them with attribution to “Joran Kikke, Agentic Atlas”.