The agent clicks into a dead-end it never saw coming
A GUI or web agent, mid-task, picks the locally-plausible next control and lands somewhere irreversible — because it never simulated where the click leads.
This is the myopia the Foresighted Planning & World Models direction forms around. Why Reasoning Fails to Plan names step-wise reasoning as step-wise greedy; BALROG documents the "knowing–doing gap." The direction is early and accelerating — 199 papers, a kind-mix of 111 limitations to just 16 findings. An open frontier, which is exactly why it's worth a bet.
Baseline today: teams paper over myopia with retries, guardrail heuristics, and human-in-the-loop babysitting — none of which give the agent actual foresight.
Observe, imagine a few futures, then commit
A lightweight world model — an LLM prompted (or cheaply tuned) to predict "if I take action X, the state becomes Y" — feeds a shallow lookahead. Roll each candidate out one or two steps, score the imagined outcomes against the goal, and only then take the best real action. No forest, no test-time weight updates.
The bet, stated so it can be lost: shallow simulated foresight beats reactive ReAct on a real, non-sandbox GUI benchmark at tolerable token cost. The idea is literature-proven (WebDreamer, Dyna-Mind); its in-the-wild reliability is not. That gap is the spike.
Where this bet could quietly eat the cycle
World-model accuracy spiral
patchPredict only a coarse textual state-delta ("a login modal appears"), enough to rank actions — never pixels. If coarse prediction can't separate good from bad, the spike has found its answer and the bet dies cheap.
Cost blowup from branching
patchHard-cap depth ≤2 and branch factor ≤4. Tokens-per-task is a first-class result, measured, not an afterthought.
Benchmark theater
patchRequire a live-ish, non-sandbox GUI/web task set — the exact regime the literature says is unsolved. A clean sandbox would flatter the method and dodge the real risk.
Explicitly out of bounds
- No test-time training or weight updates — a different, bigger bet (adjacent: Value Functions as Terminal Rewards).
- No MCTS or deep tree search — depth ≤2 only; deliberate search is a later cycle.
- No product surface — no UI, no framework integration, no packaging. Output is a measured yes/no, not a feature.
- No new infra — reuse an existing agent harness and an off-the-shelf benchmark; building either is a separate pitch.