A complete LLM agent in ~30 lines of TypeScript. An LLM in a loop, with one tool. That's it. That's an agent.

Inspired by Thomas Ptacek's You Should Write An Agent on the Fly.io blog — go read it, it's great.

Open main.ts and hit Run. The agent will fetch the Hacker News front page and tell you the top story:

tool: fetch_url({"url":"https://news.ycombinator.com/"})

>>> The top story on Hacker News right now is titled "..."

You didn't write a loop to fetch URLs. You gave the model permission to fetch, and it figured out the rest.

Rendering mermaid diagram...

The whole thing:

1. Build a messages array (the "context window" — just a list of strings).
2. Call the LLM with the messages and a list of tools.
3. If the response has tool_calls, run them, append the results to messages, loop.
4. Otherwise, print the answer and exit.

The LLM is stateless. The conversation is an illusion we cast on ourselves by replaying the whole transcript every turn.

• Add a tool. Append another entry to the tools array and a branch in run(). Try read_sqlite, send_email, roll_dice, whatever.
• Make it interactive. Wrap the loop in a for await over Deno.stdin lines (or turn it into an HTTP val with a chat UI).
• Give it memory. Persist messages in std/sqlite between runs.
• Sub-agents. Spawn another messages array with different tools. Have them talk.
• Swap the model. Change gpt-4o-mini to anything else std/openai supports.

Your wackiest idea will probably (1) work and (2) take 30 minutes to code. — tqbf

All credit for the framing and most of the prose-worth ideas goes to Thomas Ptacek (@tqbf) and the Fly.io blog: https://fly.io/blog/everyone-write-an-agent/. This val is just a Val Town / TypeScript port of the spirit of that post.