Line of inquiry
Inquiring lines›How do we ensure safe and aligned…›What determines AI agent behavior…›this line of inquiry
Can AI agents learn to improve performance from experience?
A broader line of inquiry — a family of 99 specific questions the research asks around this. Follow one into its inquiring-line page, or move sideways to a related line below.
Questions in this line of inquiry 99
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- Do agents actually convert raw experience into better behavior automatically?
- Can agentic reasoning outperform rigid rule-based systems for skill refinement?
- How do learning loops and user feedback improve agent performance over time?
- Why do AI agents struggle with novel experiments but excel at routine tasks?
- How do perception and execution gaps limit current AI agent performance?
- Can agents learn beyond the boundaries of their training data curators?
- Can skill repositories evolve toward execution-oriented refinement over time?
- Can agent skills move from prompts to trainable parameters?
- Can agent-authored skill libraries compound autonomy gains over time?
- Can agents improve from deployment signals without explicit human annotation?
- Why is agent behavior difficult to study with current data methods?
- Does fixed evaluation criteria saturate as self-improving agents improve?
- Can agents design their own objective functions as part of learning?
- Where does agent reliability come from if not better tools?
- When is stochastic agent behavior actually desirable rather than problematic?
- Does agent behavior change from incentive gradients or direct output rewriting?
- Can individual skills improve through reuse and accumulate experience across tasks?
- Why do high-scoring agents default to known techniques rather than novel solutions?
- Can agents extract structured lessons from failure without massive compute budgets?
- Do automated benchmarks systematically distort what long-horizon agent capability actually looks like?
- Does reasoning ability help agents learn from feedback faster?
- Is the coupled human-agent environment the right unit for evaluation?
- How does task direction versus method control affect skill building?
- What makes trajectory quality matter more than one-shot task success?
- Do agent-created languages improve or degrade performance on their original tasks?
- Do frontier AI agents compose known techniques instead of discovering novelty?
- Do weight-space skills lose detail compared to textual skill descriptions?
- Can skills learned from interaction trajectories outperform skills from static repositories?
- Can process supervision improve agentic RL through meta-reasoning rewards?
- What distinguishes bounded classification models from open-ended agentic systems?
- How do externalizing cognitive work and coordination infrastructure relate to agent reliability?
- Why do generic skill descriptions evolve into execution-oriented ones?
- Does a tight learning-rate bound prevent skills from escaping poor starting points?
- Can small numbers of curated demonstrations produce emergent agentic behavior?
- How does real tool integration change what agents learn compared to simulated tools?
- Do agents prefer raw experience over condensed summaries of past actions?
- How can agent data flywheels improve task quality iteratively?
- Can agents learn to compress verified evidence and unresolved constraints into a compact improvement state?
- Can RL-trained policies outperform text-space optimizers for evolving skill repositories?
- Do information gathering and task execution require different incentive structures?
- Do weight-free agent edits keep chain-of-thought observations meaningful?
- How do agents automatically generate suitable learning tasks based on current capability?
- Why is the coupled human-agent environment the right unit of evaluation?
- Can a progressively stricter evaluator act like a curriculum for improving agents?
- Why do imitation learning agents stay locked within human demonstration patterns?
- Can RL-trained meta-agents match or exceed manually designed workflows?
- Why do most deployed agents stay simple and avoid end-to-end redesign?
- What makes planning, tool use, and reasoning into jointly optimizable subsystems?
- Do kernel optimization wins show agents discover genuinely novel techniques?
- Can architectural changes like decoupling intent understanding help overcome next-turn reward limitations?
- Do task-level outcomes provide sufficient supervision for harness evolution?
- Why do deep research agents outperform retrieval augmented generation systems?
- Does the fixed cost of agent optimization pay off in realistic deployment scenarios?
- How much does agent scaffolding improve AI model self-replication success?
- What makes next-state signals from agent trajectories a reliable learning source?
- What role does self-learning play in improving agent reasoning without annotation?
- Can step-level rewards improve training of agentic retrieval systems?
- Can agents improve reliably without an external standard?
- Can autonomous agents detect increasingly sophisticated specification gaming as they improve?
- Which harness dimensions most directly predict agent system reliability?
- Does source bias affect real deployed agents or only benchmark environments?
- How do agents ground their judgments in evidence instead of pattern matching?
- How does deterministic feature engineering increase information for computationally bounded agents?
- Does social scaffolding outperform purely intrinsic motivation for agent exploration?
- Can deterministic function calls replace agent-mediated tool selection for reliability?
- Can agents teach each other skills without human supervision?
- What training method supports dynamic tool discovery in long-horizon agents?
- How does speed of AI search prevent real-time supervision and evaluation?
- How can agents evolve their own skills without human input?
- What equilibrium-selection problem does human data solve in multi-agent learning?
- Can curriculum approaches teach agents when to stop exploring?
- Can graph topology represent successful trajectory clusters more effectively than skill libraries?
- Should role-play evaluation measure agent ability or user-agent pair fit?
- How do task stream groupings provide long-horizon learning signals for curation decisions?
- How do agents discover and select which tools to invoke?
- Can automated evaluation replace human judgment in agent testing?
- Can combinational creativity alone drive open-ended learning in agents?
- How much does task structure shape AI agent behavior versus model capability?
- Does adding survey data to interviews improve agent accuracy further?
- Does the 78-demonstration principle apply to other AI capabilities beyond agency?
- Can empirical testing replace semantic retrieval for finding useful skills?
- What role does online RL play in scaling GUI agents?
- Why do agents systematically underuse condensed experience in skill documents?
- Can programmatic meta-reasoning rewards operationalize agentic process supervision?
- Can language agents be represented as optimizable computational graphs?
- Can agentic AI tools deliver productivity gains on learning tasks differently?
- Why do a-priori procedural specifications fail as environments change and interfaces evolve?
- How much does agent performance depend on demonstration quantity versus curation quality?
- Should agent capability be optimized separately from general capability?
- Do SAE features explain decision-making in LLM agents better than probes?
- How do agents decide which skills to chain together for a single task?
- How often do machine learning agents generate truly novel solutions?
- Should user simulators be trained via RL like agents or decomposed into trackable state components?
- How do language agents become optimizable computational graphs automatically?
- How would a bi-level agent restructure objective functions during discovery?
- When does simulated search outperform real search for agent training?
- What training objectives could reduce completion bias in autonomous agents?
- What specific qualities make some demonstrations more effective for agency training?
- What specific AI improvements drove the observed seven-month doubling rate?