LATCH: latency-bounded latent lookahead for constraint-safe web agents

Abstract Web agents operate under strict interaction budgets and validators, where a single invalid tool call can waste a step and an irreversible action can end an episode. Existing token-space deliberation is expensive and can hallucinate long-horizon interface states, while retrieval-only approaches do not model action-conditioned constraint propagation. We introduce LATCH, a latency-bounded latent lookahead system that makes simulate-then-select decision making practical for tool-mediated web interaction. LATCH compresses open action spaces into schema-valid candidates via an Action Proposal Module, predicts constraint-relevant state updates with a hybrid SSM+attention latent dynamics model, and stabilizes long-horizon facts through typed entity-centric memory. A calibrated critic scores short-horizon latent rollouts, enabling counterfactual ranking without generating long text futures. On WebShop, Mind2Web and WebArena under matched step budgets and shared validators, LATCH improves task success by 3.6–6.4 points over GPT-4o+ReAct while reducing constraint violations by 33–42% relative. Under a controlled decision study with a fixed candidate set, latent rollout planning yields a 12.8-point gain over token-matched LLM reranking. We further report per-decision compute (tokens, latent evaluations, and wall-time), site-holdout generalization, and budget-matched wall-time success to make the cost–reliability trade-off auditable.