A new AI system taught itself to win a classic real-time strategy (RTS) game by studying its own past matches.
Researchers built STRATA to play Red Alert by splitting command decisions across three specialized agents: one handles high-level strategy, one handles logistics, and one handles fast tactical calls. After each match, a fourth Review Agent mines the game trace for useful lessons, checks those lessons against evidence from later matches, and compresses the validated ones into short "experience cards" the strategy agent can pull from next time. In one fixed test scenario, using the learned experience cards lifted the system's win rate from 30% to 100%. Run sequentially against AI opponents with different play styles, STRATA also developed distinct long-term strategies tailored to each.
Earlier LLM-based RTS systems leaned on manually written, experience-based prompts and could miss fast-moving tactical events while waiting on slow model inference. STRATA's three-agent split handles time-sensitive combat separately from slower strategic planning, and its review-and-compress loop replaces hand-tuned prompts with evidence pulled from the system's own games.
It's still a lab result in a decades-old game, not a battlefield-ready product, but the pattern of specializing, reviewing, compressing, and repeating looks like an early sketch of AI agents that keep improving after deployment without constant human prompt-tuning.