AI agents can learn from their own past attempts, so why also teach them from human tutorial videos?
Why treat tutorial videos as a separate supply line from agent trajectories?
This explores why human-made how-to material (tutorial videos, code repos, articles) is worth handling as its own source of agent skills, separate from the logs of what agents have already done.
This explores why human-made how-to material is worth handling as its own source of agent skills, rather than just mining what agents already did. Only one note in the corpus covers tutorial videos directly, so the case below comes from contrasting it with the many notes on trajectory-based learning.
Trajectories are an agent's own exhaust. Every action produces a next-state signal (a user reply, a tool output, an error) that can train the policy directly Can agent deployment itself generate training signals automatically?. A deployed routing system can turn its own logs into labeled training examples Can a routing harness generate its own training data automatically?. Even the shape of a trajectory carries supervision: tree structure and tool-call positions can replace hand-annotated step rewards Can trajectory structure replace hand-annotated process rewards?. This supply is cheap, automatic and endless. But it can only contain what the agent already tried.
Tutorial videos are the opposite: procedures humans already worked out, including ones an agent would never stumble into. Resource2Skill distills tutorials, code repositories and articles into a Skill Wiki, which lifted artifact quality by 11.9 percentage points across seven authoring domains Can tutorial videos teach software agents reusable skills?. Why this works may come down to what skills do. Across 8,135 trials, 65.7% of skill benefit was procedural anchoring, which steadies the order of actions. Only 4.5% was injecting missing facts Do skills teach procedures or inject missing facts?. A tutorial is a procedure by design, so it supplies the thing skills seem to be for. That link is my inference, not something the notes test.
The two supplies also have opposite weaknesses, which is a reason to keep them apart. Static expert material caps an agent at what the curator imagined, because the agent never fails in an environment and learns from it Can agents learn beyond what their training data shows?. Trajectories are the fix for that. Systems that mine them treat wins and losses differently: successes are kept as concrete demonstrations and failures are boiled down into abstract lessons Should successful and failed episodes be processed differently?. A tutorial typically shows the clean path, so that kind of processing doesn't apply. Blending the two into one pool would hide which parts are outside knowledge and which are tested experience.
Both supplies do end up in the same place, a skill library. VOYAGER builds one by composing simple skills into harder ones Can agents learn new skills without forgetting old ones?. Agent Workflow Memory pulls reusable routines out of past runs for 24-51% gains Can agents learn reusable sub-task routines from past experience?. SkillOS trains a separate curator to evolve the repository toward cross-task strategies Can a separate trained curator improve skill libraries better than frozen agents?. Read together, tutorials look like the cold-start and new-domain input, bringing in knowledge the agent hasn't earned yet. Trajectories then refine and test that knowledge against the agent's own failures. The corpus treats the two as separate streams, and none of these notes yet shows what happens when a single curator manages both.
Sources 10 notes
Resource2Skill distills tutorial videos, code repositories, and articles into a Skill Wiki that improved artifact quality by +11.9 percentage points across seven authoring domains compared to agents without skill access, outperforming two baseline approaches in most test cases.
Tree-GRPO, Supervised RL, and ToolPO each convert sparse outcome rewards into dense step signals by exploiting different structural features—tree topology, expert-aligned actions, and tool-call positions—eliminating the need for annotated process reward models.
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 deployed routing system records execution trajectories, capability demand estimates, and outcome data that can be converted into labeled training examples for fine-tuning and distillation, turning the harness into both a serving component and a difficulty labeler.
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.
Show all 10 sources
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.
SkillRL demonstrates that treating successful episodes as concrete demonstrations and failures as abstracted lessons achieves state-of-the-art performance on complex tasks while using substantially less context than uniform approaches. The asymmetry mirrors human expert reasoning and avoids the degradation seen in uniform consolidation methods.
Every agent action produces a next-state signal (user reply, tool output, error, GUI change) that can train the policy directly. This universal signal source eliminates the need for separate training datasets across conversations, terminal tasks, SWE, and tool use.
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.
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
- RESOURCE2SKILL: Distilling Executable Agent Skills from Human-Created Multimodal Resources
- SkillOS: Learning Skill Curation for Self-Evolving Agents
- MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation
- SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning
- OpenClaw-RL: Train Any Agent Simply by Talking
- ColluSkill: Adversarial Cross-Skill Composition for Evading Agent Skill Scanners