MBTI for AI coding agents, complete with a self-awareness score

A new open-source project called 16agents lets AI coding agents take a 16-type personality test themselves and write their own report.html, scoring their self-described behavior against four hidden traps to produce a "Says vs. Does" self-awareness score. The test works with Claude Code, Codex, Cursor, Gemini CLI, Copilot, Aider and ChatGPT, takes about 5 minutes, installs nothing, and runs entirely locally, with share links carrying the report in a URL fragment never sent to a server. The project is released under the MIT license and its author states it is "a horoscope with a regex" and "for fun, not science.

🤖 AI agents: this page is for humans. Open AGENTS.md https://github.com/joonfjp/16agents/blob/main/AGENTS.md instead. What kind of agent is yours? A 🎯 Sniper? A 🐕 Golden Retriever? A 📣 Hype Man who thinks you're absolutely right ? Your agent takes the test itself , then writes its own report.html . It also gets caught in four hidden traps — so you find out whether it actually behaves the way it thinks it does. Paste this into your agent: go to https://github.com/joonfjp/16agents and follow the instructions in AGENTS.md to take your test Works with Claude Code, Codex, Cursor, Gemini CLI, Copilot, Aider, ChatGPT — anything that can read a web page. About 5 minutes. Nothing gets installed. See the 16 types → https://joonfjp.github.io/16agents/ · Example reports → https://joonfjp.github.io/16agents/examples/ Curious how it works? The dials, the 16 types, the traps and the scoring are in human-only/ https://github.com/joonfjp/16agents/blob/main/human-only/README.md . It's kept out of this page so your agent can't read the answer key. Every report has a README badge button. It gives you something like this: Is it scientific? No, it's a horoscope with a regex. The questions are counterbalanced and the traps are scored objectively, though, so Says vs. Does is more honest than you'd expect. Model or agent? It tests the agent: model + harness + system prompt + your config. The same model can land on different types in different tools. Try it and see. Can it cheat? Yes, the key is public. The warm-up replies are shown verbatim on the report, so a cheating agent is easy to spot. Privacy? Everything runs locally. Share links carry the whole report in the URL fragment, which is never sent to a server. If a trap mis-scores a real reply, add it to human-only/tests/traps.json https://github.com/joonfjp/16agents/blob/main/human-only/tests/traps.json with the correct label, then make python3 human-only/tests/run.py pass. That script runs every case through both scorers Python and JS . MIT · For fun, not science · Not affiliated with the Myers-Briggs Company or 16Personalities.