Agent Night demo recap: Brian Douglas's self-healing Pokémon agent

Brian Douglas of Paper Compute presented a Pokémon-playing agent at Pokémon Agent Night in San Francisco, using the game as a testbed for self-healing agent systems that learn from recorded failures. The agent runs Pokémon Red headlessly through the PyBoy emulator at roughly 100 times real-time speed, with every turn, screenshot, and session captured as a trace. Douglas argued that raw traces become useful only when they are interpreted through labeled observations, session reflections, “dreams” that identify lessons, and “inceptions” that test those lessons through repeated experiments. His system helps diagnose failures such as misreading screen boundaries or getting stuck in a room, emphasizing that recording and reviewing evidence can be more effective than simply improving prompts. Built with tools including Kafka, Flink, DuckDB, tapes.dev, and a parallel evolution harness inspired by AlphaGo and AlphaEvolve, the project evaluates multiple agent variants and reasoning levels to balance outcomes, cost, and speed. Douglas has publicly released telemetry data, code, milestones, and a blog post, presenting the project as an example of how agent memory should remain grounded in traceable evidence rather than unsupported model-generated conclusions.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.