Self-evolution is a commodity; convergence is the unsolved differentiator

Parent: recursive-harness

Hermes's "self-evolution" isn't special. The mechanism — complete a task → distill a reusable skill / reflect on the trace → reuse next time — is now table stakes, not novel.

Everyone has a flavor (lineage: Voyager 2023 → everywhere)

Voyager (2023) pioneered the skill-library-building agent.
Claude Code — CLAUDE.md memory + Skills + /remember;
Cursor / Windsurf — accumulating memories / rules;
MemGPT / Letta, Mem0, Zep — self-editing agent memory;
Reflexion — self-reflection from feedback;
ChatGPT memory — persistent user memory;
DSPy / GEPA — the prompt-optimization technique Hermes uses (not Hermes-exclusive);
plus many open-source agents (Goose, OpenHands, Cline, …).

Hermes's mild distinctiveness is packaging, not mechanism

Model-agnostic harness (skills/memory survive model swaps), open-source + self-hostable, cross-platform, GEPA specifically. Real, but the self-evolution itself is not new (and part of the hype was astroturfed).

Why "everyone has it" and "none is mind-blowing" are the same fact

They all have the easy half (the loop: distill skills, reflect on traces); none has the hard half (convergence).

Easy mechanism → everyone has it. Convergence uniformly unsolved → none is mind-blowing. They all hit the three ceilings of this thread: observability (convergence-needs-an-observable-target), deductive tractability (deducible-not-known), and unsound-loop drift (drift-is-world-wandering).

The real question isn't "who self-evolves" (all) — it's "whose self-evolution converges"

Per drift-is-world-wandering: it converges iff the world is pinned (objective metric / mechanical gate); GEPA-on-benchmarks converges, fuzzy self-eval drifts. Test for any "self-evolving" tool: does it learn against a pinned world / objective signal, or does it judge itself? The former truly improves; the latter feels good and drifts. Almost no one crosses this line — which is exactly why the feature is everywhere and the results are mediocre.