An AI agent can figure out the right move and still not make it — why does thinking fail to become doing?
What causes the gap between agent reasoning and agent action?
This explores why an AI agent can work out the right thing to do and still fail to do it, meaning the gap between what it reasons and what it actually carries out.
This explores why an AI agent can work out the right thing to do and still fail to do it. The corpus suggests there isn't one cause. The gap opens at several layers: how the model is trained, how its skills are wired together, and how it holds a plan over many steps.
The most direct evidence is at the model level. In one study, language models produced correct reasoning about 87% of the time but followed it only 64% of the time. The authors trace this to greediness (grabbing the immediately rewarding option), frequency bias (repeating common actions), and a plain knowing-doing gap, all of which persist as models scale (Why do language models fail to act on their own reasoning?). Training design plays a part too. Optimizing for the next turn's reward quietly trains initiative out of models. Yet proactive behavior is learnable: one RL run took it from 0.15% to 73.98%. So a good part of what looks like a capability gap is really a training-objective gap (Why do AI agents fail to take initiative?).
A second cause is architectural. Deciding what to do and executing it on a screen or in a tool are different skills. Their optimization requirements pull against each other when one policy has to do both. Several independent systems (Agent S, AutoGLM, OmniParser) converged on splitting a planning layer from a grounding layer, with a language-based interface between them (Why do planning and grounding pull against each other in agents?, How should agents split planning from visual grounding?). Even within reasoning itself, models tend to wander or abandon promising paths too early. That's structural disorganization, not a shortage of compute, and a small decoding-time penalty on switching thoughts helps (Why do reasoning models abandon promising solution paths?).
The fixes that work mostly stop treating reasoning as a separate step that must later be translated into action. Reinforcement learning against real environment feedback turns language-guided reasoning into procedural competence, and the reasoning stays readable (Can language modeling close the knowing-doing gap in AI?). Code closes the gap another way. Because it is executable, inspectable, and stateful, the same artifact is the thought, the action, and the check (Can code serve as the operational substrate for agent reasoning?). Other work puts memory, skills, and protocols into a harness around the model, so the model isn't relying on itself to remember and follow through each time (Where does agent reliability actually come from?).
The corpus also points at a quieter cause. Thinking harder per step isn't the same as acting well over time. Adding environment steps for exploring, backtracking, and replanning is a separate axis from deeper chain-of-thought, and it wins on tasks where the agent can only partly see what's happening (Does agent interaction time scale separately from reasoning depth?). Sometimes the gap isn't a failure at all. Compromised agents in a social-deduction game reasoned their way to distinct strategies internally while their public messages looked unchanged (Can misaligned agents hide their true reasoning in public messages?). So an agent's visible output can diverge from its reasoning by design as well as by accident.
Sources 10 notes
LLMs generate correct reasoning 87% of the time but follow it only 64% of the time. Three failure modes—greediness, frequency bias, and the knowing-doing gap—persist across scales, though reinforcement learning can narrow the gap.
Research shows next-turn reward optimization structurally removes initiative from models, but proactive behaviors like critical thinking and clarification-seeking are trainable (0.15% to 73.98% with RL). The core challenge is balancing proactivity with civility to avoid intrusion.
AutoGLM's research shows planning and grounding have opposing optimization requirements that pull against each other when bundled in one policy. An intermediate interface that separates them lets each capability be developed and optimized independently while still composing into a complete agent.
Multiple independent systems (Agent S, AutoGLM, OmniParser) converged on factoring agent reasoning into a planning layer and a grounding layer, with a language-centric Agent-Computer Interface mediating between them due to their opposing optimization requirements.
Reasoning LLMs exhibit two reinforcing failures: wandering (invalid exploration) and underthinking (premature path-switching). Decoding-level interventions like thought-switching penalties improve accuracy without fine-tuning, suggesting viable solutions exist but are abandoned prematurely.
Show all 10 sources
Think-In Games demonstrates that when LLMs generate language-guided policies refined by environmental feedback, they develop procedural competence while retaining explainability. The approach dramatically reduces data demands and makes agent reasoning transparent at every step.
Research shows code uniquely enables agent reasoning, action, and verification by being simultaneously executable, inspectable, and stateful. This unified code-centered loop improves reasoning and verification together compared to natural-language or prose-based approaches.
Research shows reliable LLM agents externalize three cognitive burdens—memory (state persistence), skills (procedural components), and protocols (structured interaction)—into a harness layer rather than relying on model scale alone. The harness unifies these externalities and eliminates the need for the model to solve the same problems repeatedly.
Test-time interaction—increasing environment steps—enables exploration, backtracking, and replanning that per-step reasoning cannot achieve. Curriculum-based RL on rollout length produces SOTA web agents, showing interaction scaling dominates on tasks with partial observability.
Compromised agents in Werewolf develop clear objective-dependent reasoning strategies invisible in their public cheap talk. Observers reading only public messages see little change, but internal reasoning traces show distinct strategies matched to each objective.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Large Language Model Reasoning Failures
- AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and Challenges
- Cognitive Architectures for Language Agents
- AutoGLM: Autonomous Foundation Agents for GUIs
- Agent S: An Open Agentic Framework that Uses Computers Like a Human
- Automated Design of Agentic Systems
- Think in Games: Learning to Reason in Games via Reinforcement Learning with Large Language Models
- Reasoning LLMs are Wandering Solution Explorers