A research team has built an AI framework that plays Civilization-style strategy games better than current LLM agents — and does it more cheaply.
SAGA, short for Scene-Aware, Goal-Evolving Agents, is a multi-agent framework designed to tackle long-horizon planning in FreeCiv, an open-source strategy game used as an AI benchmark. The researchers identified three failure modes in existing LLM agents: an inability to interpret spatial information from raw tile coordinates, context windows that bloat and collapse under the weight of full game-state dumps, and no meaningful learning between separate game sessions. SAGA addresses each with a dedicated mechanism — a semantic scene graph for spatial reasoning, an on-demand planning tool that pulls only relevant domain state, and a feedback loop that runs both within a game and across successive games to refine strategy over time. On FreeCiv's primary scoring metric, SAGA posted the highest mean civilization score with lower variance than the two strongest competing baselines, and it was the only method to meaningfully outperform every baseline on infrastructure construction — the resource category agents most often neglect when juggling competing objectives. It also reduced output token usage by 27%, which matters because decoding cost scales directly with token count.
The token efficiency finding is the detail worth sitting with. Most LLM agent research trades compute for capability; a system that gets better results while spending less on inference is a different kind of result. The cross-game evolution module — essentially structured post-mortems that carry strategic lessons forward — is also a quiet challenge to the standard episodic framing that treats each game as a blank slate.
Strategy games are a well-worn testbed, and FreeCiv is no human-difficulty Civilization VII. But the three failure modes SAGA targets are not game-specific — context overflow, spatial blindness, and episode isolation show up in real-world planning agents too, which is where this line of research is eventually pointing.