When an AI agent learns from its own mistakes, does the improvement live in the model itself, or somewhere else?
How can agent data flywheels improve task quality iteratively?
This explores how an agent's own work (its attempts, failures and feedback) can be recycled into better performance on the next round, and where that improvement actually gets stored.
This explores how an agent's own work (its attempts, failures and feedback) can be recycled into better performance on the next round, and where that improvement gets stored. None of these notes use the phrase 'data flywheel', but together they cover its parts. The clearest way in is to look at what a flywheel is not. Agents trained on a fixed set of expert demonstrations never interact with an environment during training, so they can't learn from their own failures. Their ceiling is set by what the dataset's curators imagined, not by what the agent could do Can agents learn beyond what their training data shows?. A flywheel breaks that ceiling by making the agent's own experience the data.
The surprise is that the loop often doesn't touch the model's weights. A survey of self-improving agents splits them into a slow loop that updates model weights and a fast loop that updates prompts, memory and tools. Most recent progress is in the fast loop, because scaffold changes are cheaper and reversible Do self-improving agents really split into two distinct loops?. AgentFly is one example. It keeps memories of past cases, subtasks and tool use, and improves through those memory operations alone, reaching 87.88% on the GAIA validation set with the LLM unchanged Can agents learn continuously from experience without updating weights?. Voyager stores what worked as executable skills, builds harder skills out of simpler ones, and uses an automatic curriculum to pick what to try next. Each round starts from a bigger toolbox, and nothing is forgotten the way it can be with weight updates Can agents learn new skills without forgetting old ones?.
The harness around the model can also be the thing that improves. Running automated research loops across many environments found four harness mechanisms (action execution, context compaction, observation handling and delegated reading). They cut token traffic by 44.7–49.0% at comparable performance Can agent harnesses be automatically optimized across many environments?. Tuning the execution system around frozen weights lifted Terminal-Bench 2.1 scores across several models, and the same runbook carried over to newer models unchanged Can execution harnesses lift model performance without retuning weights?. So a flywheel that stores its gains in the harness keeps them when the model is swapped out. The weight-updating version also exists. FlowReasoner trains a meta-agent with reinforcement learning on real execution feedback, and it designs a different multi-agent workflow for each query Can AI systems design unique multi-agent workflows per individual query?.
What keeps the wheel turning is persistence. Across 17 frontier models on long optimization tasks, the strongest predictor of success was how many benchmark-edit-incorporate cycles a model completed within its time budget, more than the quality of its first attempt. Most models quit early or spent their budget without progress What predicts success in ultra-long-horizon agent tasks?. The loop also needs somewhere to put what it learns. Reliable agents move memory, skills and interaction protocols out of the model and into the harness, so the model doesn't have to solve the same problem again each time Where does agent reliability actually come from?. The retrievals here say little about what stops a flywheel from amplifying its own mistakes, so that question is still open in this collection.
Sources 9 notes
Agents trained on static expert datasets cannot learn from their own failures or generalize beyond demonstrated scenarios because they never interact with environments during training. Competence is capped by what curators imagined, not by agent capacity.
A survey framework organizes self-improving agents into two update mechanisms: slow parametric loops updating foundation model weights, and fast non-parametric loops updating prompts, memory, and tools. Recent progress concentrates in the fast loop because scaffold updates are cheaper and reversible than weight updates.
AgentFly formalizes agent learning as a Memory-augmented MDP with three memory modules (case, subtask, tool) that enable credit assignment and policy improvement entirely through memory operations. The approach achieved 87.88% on GAIA validation without modifying LLM parameters.
VOYAGER demonstrates that storing executable skills in an embedding-indexed library and composing complex skills from simpler ones allows agents to learn continuously while avoiding the forgetting that occurs with weight-update-based methods. Environmental feedback refines skills while an automatic curriculum drives continual exploration.
Cross-environment harness optimization yielded four mechanisms (action execution, context compaction, observation handling, delegated reading) that reduced token traffic by 44.7–49.0% while maintaining comparable performance on a 51-task benchmark, suggesting harness-level gains are orthogonal to model improvements.
Show all 9 sources
StateM improves Terminal-Bench 2.1 accuracy across multiple models by optimizing execution systems around frozen weights. The same runbook transfers to newer models without modification, achieving 95.3% on GPT-5.6 and lifting DeepSeek-V4 Flash by 5.4 points.
FlowReasoner demonstrates that meta-agents trained with reinforcement learning and external execution feedback can generate unique multi-agent architectures for each user query, optimizing across performance, complexity, and efficiency—moving beyond fixed task-level workflow templates.
Across 17 frontier models on 36 expert-curated optimization tasks, repeated benchmark-edit-incorporate cycles within a wall-clock budget proved the dominant success predictor. Most models terminated early or burned budget unproductively; Claude Opus 4.6 stood out as persistent.
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.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Demystifying Agent Skills: Why They Work-Until They Don't
- Harness Updating Is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents
- Rethinking the Evaluation of Harness Evolution for Agents
- SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning
- Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments
- Useful Memories Become Faulty When Continuously Updated by LLMs
- StateM: Reaching 95.3% Raw Accuracy, or a \$15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling
- AutoLab: Can Frontier Models Solve Long-Horizon Auto Research and Engineering Tasks?