If an AI agent hits a task it was never prepared for, can it teach itself the missing skill on the job?
Can agents acquire new skills online when offline skill coverage runs out?
This explores whether an agent that ships with a fixed set of skills, written or trained before deployment, can pick up new ones while it is running when a task falls outside what that set covers.
This explores whether an agent that ships with a fixed set of skills can pick up new ones on the job when a task falls outside that set. The corpus says yes, and there is more than one way to do it. The routes differ in where the new skill gets stored and when it gets learned.
Offline coverage runs out because it is capped by what someone imagined in advance. Agents trained on static expert demonstrations never interact with an environment while training, so they can't learn from their own failures, and their competence stops at what the curators thought of Can agents learn beyond what their training data shows?. What's missing in those gaps is also not mostly facts. An analysis of 8,135 trials found skills work as procedural anchors in 65.7% of cases, versus 4.5% for injecting knowledge Do skills teach procedures or inject missing facts?. So online learning has to produce reusable procedures, not just look things up.
The most direct route is to write the skill in the middle of the task. MUSE-Autoskill makes skill creation a tool the agent calls inside its own reasoning loop. The new skill is grounded in the exact task context, gets immediate feedback, and is validated at runtime. It reaches 87.94% task accuracy and transfers to other agents with little loss Does creating skills inside the agent loop eliminate mismatches?. VOYAGER stores executable skills in an indexed library and builds complex skills from simpler ones. Environmental feedback refines them, and an automatic curriculum keeps pushing exploration into new territory Can agents learn new skills without forgetting old ones?. Agent Workflow Memory extracts sub-task routines from past experience, strips out example-specific details, and stacks routines into hierarchies. Its gains were 24.6% on Mind2Web and 51.1% on WebArena, and they grew as the gap between training and test tasks widened Can agents learn reusable sub-task routines from past experience?. That widening gap is the situation where offline coverage runs out.
Most of these approaches never touch the model's weights. AgentFly treats learning as memory operations over past cases, subtasks and tools. It reached 87.88% on GAIA validation with no parameter updates Can agents learn continuously from experience without updating weights?. MetaClaw adds a second, slower timescale. Failures turn into new skills within seconds with no downtime, and gradient-based training happens later in idle windows. The two feed each other: better policies produce more informative failures, and richer skills produce better trajectories Can agents adapt without pausing service to users?. Tools follow the same logic. DeepAgent discovers tools as it goes rather than relying on a pre-retrieved set, which matters when the tool space is too large to list up front Can agents discover tools dynamically instead of pre-selecting them?.
Two further pieces address the quality of what gets learned and who benefits. SkillOS trains a separate curator to maintain the skill repository while the executor stays frozen. The repository shifts from generic, verbose additions toward actionable execution logic and cross-task meta-strategies Can a separate trained curator improve skill libraries better than frozen agents?. SkillClaw aggregates trajectories from many users, so a gap one person hits becomes a skill everyone gets, with no manual curation How can agent systems share learned skills across users?.
Put together, online skill acquisition is mostly a job for the harness around the model, not the model itself. That fits the broader finding that reliable agents move memory, skills and protocols into external structure instead of leaning on model scale Where does agent reliability actually come from?. The corpus reports these results on different benchmarks and doesn't compare the routes head to head, so it can't yet say which one is best.
Sources 11 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.
Analysis of 8,135 trials shows procedural anchoring accounts for 65.7% of skill cases versus 4.5% for knowledge injection. Skills fail when retrieved incorrectly, invoked out of context, or followed too rigidly.
MUSE-Autoskill demonstrates that invoking skill creation from within the agent's reasoning loop grounds new skills in exact task context, immediate feedback, and runtime validation. In-loop skills reach 87.94% task accuracy and transfer to other agents with minimal loss, eliminating the situated context problem of offline authoring.
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.
Agent Workflow Memory induces sub-task routines at finer granularity than full tasks, abstracts example-specific values, and compounds them hierarchically. This produces 24.6% relative gain on Mind2Web and 51.1% on WebArena, with larger gains as train-test gaps widen.
Show all 11 sources
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.
MetaClaw demonstrates that deployed agents require both rapid skill injection from failures (seconds, zero downtime) and slower gradient-based optimization during idle windows (minutes to hours). The two mechanisms reinforce each other, with better policies producing more informative failures and richer skills enabling higher-reward trajectories.
DeepAgent demonstrates that discovering tools as needed—rather than pre-retrieving a fixed set—enables agents to maintain global task perspective and adapt strategy mid-execution. This approach scales better for long-horizon tasks where the tool space is too large to enumerate.
SkillOS shows that separating a trainable curator from a frozen executor, grouped by task streams, causes skill repositories to shift from generic verbose additions toward actionable execution logic and cross-task meta-strategies. The trained curator generalizes across different executor backbones and domains.
SkillClaw aggregates interaction trajectories across users, processes them through an autonomous evolver that identifies patterns and refines skills, then synchronizes updates system-wide. This converts siloed individual learning into shared capability improvement without manual curation.
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
- SkillClaw: Let Skills Evolve Collectively with Agentic Evolver
- SkillOS: Learning Skill Curation for Self-Evolving Agents
- MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation
- RESOURCE2SKILL: Distilling Executable Agent Skills from Human-Created Multimodal Resources
- MetaClaw: Just Talk — An Agent That Meta-Learns and Evolves in the Wild
- LatentSkill: From In-Context Textual Skills to In-Weight Latent Skills for LLM Agents
- Harness-R1: Learning to Edit Executable Runtime Harnesses from Agent Failure Trajectories