Thinking Fast, Anywhere: One System-1 for the Decision Points of Any Agent Workflow
Abstract
AnyS1 converts procedural skills into shared decision points, trains a small calibrated pointer model to resolve routine choices, and falls back to a large model when uncertain, reducing cost and latency while preserving reliability.
Autonomous LLM agents incur substantial cost and latency by invoking a large model (System-2) at every step, even when many intermediate actions choose among a small set of known candidates. Existing dual-system agents offload such decisions to a lightweight System-1 model, but rely on workflow-specific hooks authored per environment. We present AnyS1, a framework that turns human-authored procedural skills, whose instructions naturally specify where choices arise and which options apply, into a shared, domain-general source of decision points. Offline, AnyS1 mines grounded enumerable decisions across over 1,000 public skills, gathers execution-grounded trajectory states inside isolated containers to construct Any-DP, and trains a single calibrated 4B pointer System-1 with label smoothing. At run time, compiled workflows resolve routine decisions in one forward pass and fall back to System-2 when confidence falls below a calibrated threshold. Across three unseen benchmarks and distinct System-2 model families, AnyS1 preserves task reliability while cutting episode cost by 75-91% and model-call time by 63-92% relative to an uncompiled System-2 agent, and maintains performance under zero-shot threshold transfer.
Community
Thinking Fast, Anywhere: One System-1 for the Decision Points of Any Agent Workflow
Where this lands in production is the candidate list. The whole bet is that decision points are enumerable from human-authored skill docs, and the steps that cost me the most are the ones where the right option isn't in the list at all — the tool returned something nobody planned for. Escalating those to System-2 at runtime is exactly where the latency win evaporates, since you've already paid for the fast pick plus the slow rescue. So I want two numbers: the fraction of real trajectory steps that are genuinely enumerable, measured on logs rather than on the tasks the docs were written for, and what end-to-end success does when the System-1 pick is wrong. A wrong pick that keeps going with no reasoning trace to read is a worse failure mode than a slow correct step — at least the slow one leaves you something to debug.
Get this paper in your agent:
hf papers read 2611.23751 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 1
Datasets citing this paper 1
smSquirrel/SkillDP
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper