INQUIRING LINE

What if AI agents could teach each other skills directly, with no human trainer involved at any step?

Can agents teach each other skills without human supervision?

This explores whether agents can generate and pass skills to each other — or to future versions of themselves — using their own signals instead of human labels, curated demonstrations, or hand-written rewards.


This explores whether agents can generate and pass skills to each other without human supervision, and the corpus suggests the answer is increasingly yes — but the interesting part is *how* the human gets designed out of the loop. The starting problem is that agents trained on fixed human demonstrations are capped by whatever the curator imagined; they never touch an environment, so they can't learn from their own failures Can agents learn beyond what their training data shows?. To teach each other, agents first need a source of feedback that isn't a person. One route is to *manufacture* that feedback internally: a self-play loop where a Challenger raises difficulty as a curriculum and a neutral Judge hands out binary verdicts as reward lets language skills co-evolve with no human in the loop at all Can language models learn skills without human supervision?. Another route treats the agent's own actions as the teacher — future states become the supervision signal, matching demonstration-trained baselines on half the data Can agents learn from their own actions without external rewards?.

Once an agent can learn a skill, the question becomes how it hands that skill to others. Here the corpus splits into two very different transmission channels. The first is *externalized libraries*: skills stored as executable code that any agent can retrieve and compose into harder skills, which is how VOYAGER learns continuously without overwriting old abilities Can agents learn new skills without forgetting old ones?. Scale that up and you get systems that aggregate interaction trajectories from many users, mine them for patterns, and sync refined skills back out across the whole ecosystem — collective learning with no manual curation How can agent systems share learned skills across users?. Intriguingly, the librarian itself can be a *separate trained agent*: decoupling a trainable curator from a frozen executor pushes the shared repository away from verbose junk and toward reusable meta-strategies, and that curator generalizes to executor backbones it never trained on Can a separate trained curator improve skill libraries better than frozen agents?.

The second channel skips language and files entirely. Instead of writing a skill down, agents can share the internal representations behind it — passing KV caches or hidden embeddings directly, which preserves reasoning fidelity that text throws away and cuts token cost dramatically Can agents share thoughts without converting them to text?. A related line formalizes recovering individual, shared, and private latent thoughts from hidden states, so agents can even detect where their internal goals conflict before that shows up in words Can agents share thoughts directly without using language?.

The thing you might not expect: several of these systems teach without ever touching model weights. AgentFly improves policy entirely through memory operations — case, subtask, and tool memories doing credit assignment — hitting 87.88% on GAIA with the base model frozen Can agents learn continuously from experience without updating weights?. And skills learned in a multi-agent setting can transfer *out* of it: training a model to delegate subtasks and integrate the results teaches disciplined decomposition that carries over to solo tasks, so the 'teaching' isn't just orchestration, it's a portable habit of thought Can delegation teach models to manage context more actively?. Cooperation itself can emerge the same way — agents trained against diverse partners learn to best-respond into cooperation with no hardcoded rules Can agents learn cooperation by adapting to diverse partners?. The recurring caveat across all of it: unsupervised skill transfer needs a governor — a generalization safeguard against self-play collapse, or contribution scoring to prune agents that are dragging the team down Can multi-agent teams automatically remove their weakest members?. Left ungoverned, agents teaching agents can just as easily reinforce each other's bad habits.


Sources 12 notes

Can agents learn beyond what their training data shows?

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.

Can language models learn skills without human supervision?

Ctx2Skill's three-role self-play loop manufactures missing feedback through internal signals: the Challenger escalates difficulty as curriculum, the Judge gives binary verdicts as reward, and both sides evolve via natural-language skill edits. Success requires balancing adversarial pressure against a generalization safeguard to prevent collapse.

Can agents learn from their own actions without external rewards?

Research across eight environments shows that agents can use future states from their own actions as supervision without external rewards, matching expert-dependent baselines with half the data and providing superior warm-starts for subsequent RL training.

Can agents learn new skills without forgetting old ones?

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.

How can agent systems share learned skills across users?

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.

Show all 12 sources
Can a separate trained curator improve skill libraries better than frozen agents?

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.

Can agents share thoughts without converting them to text?

LatentMAS enables agents to share internal representations directly via KV caches, reaching 14.6% accuracy gains and 70.8-83.7% token reduction with no additional training. Hidden embeddings preserve reasoning fidelity that text-based systems cannot.

Can agents share thoughts directly without using language?

Research formalizes inter-agent thought sharing via sparse autoencoders that recover individual, shared, and private latent thoughts from hidden states. This approach detects alignment conflicts at the representational level before they manifest in language.

Can agents learn continuously from experience without updating weights?

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.

Can delegation teach models to manage context more actively?

SearchSwarm shows that training models to delegate subtasks and integrate summarized results beats passive compression, with a 30B model matching much larger ones. Critically, the delegation skill transfers to single-agent tasks, suggesting it teaches disciplined decomposition and evidence grounding, not just orchestration.

Can agents learn cooperation by adapting to diverse partners?

Sequence model agents trained against diverse co-players develop in-context best-response strategies that naturally resolve into cooperation. Mutual vulnerability to exploitation creates pressure that drives cooperative mutual adaptation without hardcoded assumptions or timescale separation.

Can multi-agent teams automatically remove their weakest members?

DyLAN's three-step importance scoring mechanism (propagation, aggregation, selection) quantifies individual agent contributions and automatically removes uninformative agents during inference, optimizing team composition without task-specific tuning.

Papers this line draws on 8

The research behind the notes this line reads — ranked by how closely each paper relates.

Research prompt for your LLMexpand ↓

Copy into ChatGPT or Claude to take this line of inquiry further — it asks the model to find newer work and re-test which earlier constraints still hold.

You are an AI-agents research analyst. Question, still open: can agents teach each other skills without human supervision — and how does the human get designed out of the loop?

What a curated library found — and when (dated claims, not current truth; findings span 2023–2026):
- Agents trained on fixed human demonstrations are capped by the curator's imagination; never touching an environment, they can't learn from their own failures (~2025).
- Manufactured feedback works: a Challenger-raises-curriculum + neutral-Judge self-play loop co-evolves language skills with no human present; a separate line matches demo-trained baselines on half the data using future states as supervision (~2025).
- Skills transmit two ways: externalized executable libraries composed into harder skills (VOYAGER-style), and latent channels passing KV-caches/hidden embeddings for lossless, cheap transfer (~2025–2026).
- Weights can stay frozen: AgentFly hits 87.88% on GAIA via memory operations alone; multi-agent delegation habits transfer OUT to solo tasks (~2026).
- Recurring caveat: unsupervised transfer needs a governor against self-play collapse and to prune low-contribution agents (~2026).

Anchor papers (verify; mind their dates): Agent Learning via Early Experience (2025); Thought Communication in Multiagent Collaboration (2025); SkillOS: Learning Skill Curation for Self-Evolving Agents (2026); SearchSwarm (2026).

Your task:
(1) RE-TEST EACH CONSTRAINT, biased toward what's still UNSOLVED. For every finding, judge whether newer models, training, tooling, orchestration (memory, caching, multi-agent), or evaluation have relaxed or overturned it. Separate the durable question from the perishable limit; cite what resolved each, and say plainly where a constraint (e.g. self-play collapse) still holds.
(2) Surface the strongest superseding work from the last ~6 months.
(3) Propose 2 research questions that assume the regime has moved.

Cite arXiv IDs; flag anything you cannot ground in a real paper.