A 1990s BBS game as a modern AI testbed

May 25, 2026

One of the more unexpected projects in the lab: pointing modern AI agents at TradeWars 2002, a text-based space-trading game that ran on dial-up bulletin board systems in the 1990s. It sounds like nostalgia. It’s actually a clean little laboratory for agent behavior.

Why an ancient game?

TradeWars has exactly the properties you want in a testbed:

• A constrained, text-only interface — no vision model required, just reading and typing.
• Clear goals and a score — trade, expand, compete; success is measurable.
• Real strategy — resource management, risk, and adversaries make it non-trivial.
• A self-contained world — you can run a private universe and let agents loose safely.

That combination makes it ideal for testing how an agent plans, remembers, and adapts over a long session — without the cost or unpredictability of turning it loose on the open internet.

What I’m building

The plan is a framework where an AI agent connects to a private game server, reads the game state, reasons about its next move, and acts — then an “autoresearch” loop studies how different models and prompting strategies actually perform. A separate server runs the game in an isolated, locked-down universe.

Local-first, on purpose

The interesting part: production runs are meant to happen on local AI hardware, not a cloud API. Long agent loops make a lot of calls, and keeping inference local keeps the cost predictable and the experiment fully under my control.

Starter draft — expand with architecture, the autoresearch loop, and early results.

← all posts