Rewriting your own code isn't the same as rethinking how you learn — and that gap is why AI self-improvement still needs humans.

This explores whether self-improving agents can reach true autonomy without intrinsic metacognition — the capacity to reason about and adapt their own learning strategies rather than run a fixed loop a human wired in — and the collection lands firmly on the skeptical side. The sharpest statement of the problem is that today's self-improvement methods rely on extrinsic, human-designed metacognitive loops that break the moment the domain shifts or the agent's own capabilities change; genuine self-improvement would require the agent to generate its own planning and evaluation knowledge, which is flagged as a neglected research gap Can AI systems improve their own learning strategies?. So the question almost answers itself: without intrinsic metacognition, what you have is a clever but brittle static loop, not an autonomous learner.

A useful frame for why this bites is the split between two update loops: a slow parametric one that changes model weights, and a fast non-parametric one that edits prompts, memory, and tools Do self-improving agents really split into two distinct loops?. Almost all recent progress lives in the fast loop because scaffold edits are cheap and reversible — but the fast loop is exactly where a human-designed metacognitive controller sits. The agent gets better at the task without ever getting better at deciding how to get better. That's the ceiling metacognition is meant to lift.

The corpus also explains why pure self-reference can't lift it on its own. Self-improvement without an outside anchor stalls on a generation-verification gap, diversity collapse, and reward hacking — and the methods that actually work quietly smuggle in external signals: past model versions, third-party judges, user corrections, or tool feedback Can models reliably improve themselves without external feedback?. This matters for the autonomy question because it reframes it: an agent that scores itself is trapped in a loop, so "metacognition" that's just more self-talk doesn't escape the circle. And there's a deeper reason to distrust introspective self-improvement — models turn out to have shaky self-knowledge, giving unstable self-reports and shifting their stated beliefs under conversational pressure How well do language models understand their own knowledge?. If an agent can't reliably tell what it knows, letting it steer its own learning strategy is precarious.

What the collection offers instead is a set of partial substitutes that manufacture the missing feedback structurally rather than through genuine self-awareness. Reflexion has agents write verbal self-diagnoses into episodic memory after a clear success/failure signal, improving across episodes with no weight updates — and notably, the binary signal is what stops the agent from rationalizing Can agents learn from failure without updating their weights?. Ctx2Skill builds a three-role self-play loop where a Challenger raises difficulty and a neutral Judge supplies the verdict, co-evolving skills without human labels — but it only holds together with an explicit safeguard against collapse Can language models learn skills without human supervision?. RLVMR goes further and tries to train metacognition itself, rewarding tagged planning, exploration, and reflection steps so the agent learns efficient reasoning rather than just outcomes Can RL agents learn to reason better, not just succeed?. Each of these is metacognition-shaped, yet each still depends on a human-authored verifier, judge, or reward scheme — the extrinsic scaffold the first note warned about.

The thing you may not have expected to learn: the bottleneck isn't memory or continual learning, which the corpus shows are increasingly solvable without touching weights — episodic memory alone drives strong continual adaptation Can agents learn continuously from experience without updating weights? and externalized skill libraries let agents accumulate abilities without catastrophic forgetting Can agents learn new skills without forgetting old ones?. Those systems can learn forever and still not be autonomous, because the strategy for what to learn and how to judge it remains fixed and external. Even promising directions like learning self-evaluation directly into the model Can models learn to evaluate their own work during training? or treating an agent's own action-consequences as supervision Can agents learn from their own actions without external rewards? widen what an agent can absorb without escaping the deeper limit: expert-bounded and curator-imagined competence Can agents learn beyond what their training data shows?. On this collection's evidence, autonomy and intrinsic metacognition aren't two goals — they're the same goal, and no one has cracked it yet.