Many-Lives is a persistent-world agent harness, demonstrated by a turn-based text dungeon. The harness is the product; the game is the environment. It solves the problem that LLM "agents" usually live in a chat transcript: continuity depends on replaying history, context grows without bound, and the model is trusted to report its own effects. Many-Lives inverts that. Models interpret, propose, describe and reason; only deterministic application code may establish or mutate canonical state. No model has database-write access. Player text is an attempted action, never an authority, so prompt-injection attempts change nothing. Dice come from seeded code RNG. Narration is generated after commit and can never alter state. Repeating a turn id returns the stored result and applies nothing. Continuity comes from stored state, not a transcript: kill the server mid-raid, restart, and the same world resumes. Measured result: stored history grew 100x (100 to 10,000 events, 3.8 MB) while the context assembled for one model call moved from 239 to 246 tokens — 1.03x, against a 3000-token budget. Reproducible with scripts/seed_stress_history.py.
More from The Harness Engineering & Model Wrangling Hackathon
Ripple designs circuit boards from plain English, grades them against hidden checks, and evolves its own harness from the failures, going from 0/4 to 3/4 on…
+1Offload learns the work you repeat, so next time you can hand it over. Its engine, REM (Replay · Evolve · Merge), is an agent harness with a day/night cycle…
Recursive Computer Use is a local-first desktop automation harness that enables AI agents to execute tasks on your computer, verify outcomes locally, and…
Da Vinci is a recursive improvement CAD harness designed to support physical engineering tasks through specialized AI agents, programmatic CAD and automated…
CyclopsDiary gives factory workers wearing Meta AI glasses a shared memory of the floor. Every worker's glasses feed one MongoDB memory that keeps what was…
A harness that teaches itself what context to show Jev, TypeSafe's fast decision model, proves the result on a storm it never saw, and can't cheat because…
