Long-task AI Agent success rates have long hovered around 9.85%. At ICML 2025, the SqueezeAILab team pushed that to 57.58% — with one core move: splitting the agent into two roles. An Agent, in this context, is an AI program that autonomously executes multi-step tasks — say, clicking through a dozen screens on a website to get something done.
What this is
PLAN-AND-ACT's logic isn't complicated. When agents blow up on long tasks, it's usually not because the model is too dumb — it's because one model is being crushed by two jobs at once: planning the overall strategy and grinding through each concrete action. The paper separates these two jobs completely. A PLANNER draws the route; an EXECUTOR follows it. After every step, the PLANNER replans based on the current state instead of stubbornly clinging to the old plan.
The numbers are the hard evidence. On the WebArena-Lite benchmark, success rate jumped from a 9.85% baseline to 57.58% — 8.48 points above the previous SOTA. On WebVoyager's text-mode tasks, it hit 81.36%. The 70B PLANNER and EXECUTOR models, training code, and synthetic data pipeline are all open-sourced; the pipeline can crank out 15,000 training samples per hour.
Industry view
Supporters read this as a "subtraction victory" for Agent engineering. The field had defaulted to assuming failure meant models weren't big enough or context windows weren't long enough; this paper uses numbers to show the real problem is role allocation. The open-sourced models and reproducible synthetic data pipeline also lower the bar for training a planning model.
But the counter-arguments deserve a line in the ledger. First, 57% success is still a long way from production-grade — real business workflows typically demand 90%+, and right now this looks more like "demo that runs." Second, 70B models aren't cheap on inference compute; mid-sized teams may not be able to absorb the cost. Third, it bets on a "heavy retraining" path — train your own planner — which runs counter to the prevailing RAG and Function Calling (letting the model directly retrieve from a knowledge base or call external tools) "add a layer and ship" direction. Which bet to make depends on your team's resources.
Impact on regular people
For enterprise IT: Long-horizon Agents have real deployment potential in multi-step workflows like customer service and e-commerce operations — but 57% success means human-in-the-loop fallback is still required. You can't replace people outright.
For office workers: The first change white-collar workers will feel probably isn't Agents doing your job for you — it's Agents doing half of it. AI runs the first stretch; when something breaks down the back half, a human takes over. The boundary between "operations" and "judgment" inside a role gets redrawn.
For consumer markets: Everyday users won't see direct benefit in the short term. The paper's demos are all web-scale operations, still a meaningful distance from "tap a few times on your phone and let AI book dinner or a flight." Most of the value lands in B2B first.