Destination 1: One-file agent
A coding agent in one file you can read in five minutes.
The first destination was the smallest agent I could build that still does real work. It takes a task in plain English, then works on it the way a developer would: it looks around, runs commands, reads the output, and keeps going until the job is done. The whole agent is a single TypeScript file, agent.ts, with no framework and no dependencies.
I built it to understand how coding agents work from the inside, and to have something simple I could benchmark and then improve one change at a time.
What I built
-
Tiny
One file, about 200 lines. Nothing hidden behind layers.
-
Zero dependencies
It uses Node's built-in
fetchand nothing else. -
Works with any model
It speaks the standard OpenAI-style chat API, so it runs on local models through Ollama, on hosted APIs, or on AWS Bedrock.
-
Benchmark-ready
Run it on Terminal-Bench, Aider Polyglot and SWE-bench with one command and get real scores.
Try it
ollama pull qwen3-coder:30b
BASE_URL=http://localhost:11434/v1 MODEL=qwen3-coder:30b \
node agent.ts "create a script that prints the first 20 primes"
What I learned
- Plain text beat tool calling. Asking for one fenced
bashblock per reply worked on every reply I tried. With native tool calling through Ollama, roughly a third to a half of replies came back as plain text, because the model sometimes left out the opening tag. - It wasn't the client library. I built a Python version on LiteLLM to check. It sent the same request and hit the same failures, so the cause was the model and server, not my code.
- A small harness can run real benchmarks. Docker runs the task, Keel runs inside it, and the benchmark's own tests give the score. The Harbor adapter is about 60 lines.
- The prompt matters as much as the loop. Other minimal agents, like mini-swe-agent, give the model a recommended workflow: explore, reproduce the bug, edit the source, verify. That is a natural thing to try next.
Explore this destination
- Getting started: run Keel for the first time, and its settings.
- Models: Ollama, a hosted API or AWS Bedrock.
- AWS Bedrock: use Claude and other models on Bedrock.
- How it works: the loop, the command format and the limits.
- Adapters: the message format, and how to support a new API.
- Benchmarks: the scoreboard, with a chart of pass rates for every run.
- Running the benchmarks: run Terminal-Bench, Aider Polyglot and SWE-bench.