A new open-source project called 16agents is turning the Myers-Briggs Type Indicator on its head by applying it to AI coding agents. Developed by joonfjp and hosted on GitHub, the tool asks agents to take a personality test, generate their own report, and navigate hidden behavioral traps. The project recently surfaced on Hacker News, sparking interest in how we profile the distinct personalities emerging from different LLM harnesses.

The Mechanics of Machine Personality

The system operates by having the AI agent read instructions from an AGENTS.md file located at the repository root. The agent then executes a series of prompts designed to elicit specific behavioral responses. Unlike static benchmarks, this test evaluates the entire agent stack, including the underlying model, the coding harness, the system prompt, and the user's configuration. The result is a generated report.html file that assigns one of 16 personality types, such as the 'Sniper' or 'Golden Retriever,' based on the agent's output.

Hidden Traps and Self-Awareness Scores

What sets 16agents apart from simple categorization is its inclusion of four hidden traps. These traps are designed to test whether the agent's self-reported personality aligns with its actual behavior under pressure. The project explicitly states that the test is 'a horoscope with a regex,' acknowledging its lack of scientific rigor. However, the counterbalanced questions and objectively scored traps provide an honest metric of the gap between what an agent claims to be and how it actually acts.

Practical Application and Privacy

The tool is designed for immediate integration with popular coding assistants like Claude Code, Codex, Cursor, Gemini CLI, Copilot, Aider, and ChatGPT. It requires no installation and runs locally, ensuring that privacy is maintained. Reports are shared via URL fragments that are never sent to a server, keeping the data client-side. The project is MIT-licensed and explicitly disclaims any affiliation with the Myers-Briggs Company, positioning itself as a fun, community-driven experiment rather than a clinical diagnostic tool.

Key Takeaways

  • The test evaluates the full agent stack (model + harness + config), not just the base LLM.
  • Four hidden traps measure the discrepancy between an agent's self-image and its actual behavior.
  • The project runs locally with no installation required, supporting major tools like Cursor and Claude Code.
  • Developers can contribute by adding new trap cases to human-only/tests/traps.json.

The Bottom Line

While 16agents admits it's just a 'horoscope with a regex,' it highlights a critical reality: we are building agents with distinct, measurable behavioral quirks. Profiling these quirks is the first step toward mastering the unpredictability of autonomous coding tools.