Self-improving agents

Agents that author and refine their own capabilities — writing new agent-skills from experience and improving them through use, rather than waiting for a human to update them. The "growth" axis of the agent landscape: capability that compounds from the agent's own runs.

Distinct from neighbours

vs. durable-agents — durability is persisting state across time; self-improvement is changing the agent's own skill/knowledge set over time. Often paired (you need persistence to accumulate improvements) but not the same.
vs. agent-skills — skills are the format; this is the agent generating and editing them autonomously.

Instances

hermes-agent — a "closed learning loop": autonomously creates skills after complex tasks and refines them in use; agent-curated memory that builds "a deepening model of you." hermes-learn-skill-capture makes the skill-authoring side concrete and names an autonomy spectrum: hand-write (you) → /learn (agent authors a SKILL.md from a doc/URL/past conversation you point it at) → skill_manage auto (the agent captures its own procedural memory after a 5+ tool-call workflow, unprompted) → Hub install (community). A write_approval gate can route agent-written skills to a review queue — though it's off by default (the agent writes skills independently), which is exactly the drift/quality knob the open question below turns on.
adk — "self-generating meta-skills" (an agent that writes its own skills) adk-agents-with-skills.
gstack — the Reflect step of its sprint codifies learnings for reuse.
zouroboros — daily introspection cycles (a "Health Council") audit capability gaps, prescribe fixes, and evolve procedures — but gate each procedure change through a three-model consensus vote, a concrete answer to the drift/quality risk below.
autoharness — moves up from authoring one skill to managing the whole library: distils Claude Code sessions into skills, consolidates similar ones instead of accumulating duplicates, and archives by invocation rate. Its drift control is scoping (it only edits skills it authored) + pre-deploy validation, with real usage/adherence as the feedback signal — a fourth design point on the drift question alongside zouroboros's vote, autonovel's LLM-judge, and autoresearch's objective metric.
auto-browser — an experimental "Convergence harness" (Stage 0): a browser-agent tool records a tamper-checked trace of a completed task and writes a staged skill candidate with signed provenance attached; unsigned candidates are refused and mock-run candidates are tagged simulated. Promotion out of the staging area stays a manual, explicit step, which puts its drift control closer to zouroboros's consensus gate below than to Hermes's off-by-default review queue.
omega-framework — the negative instance (2026-09-20). Its disclaimer says the agent may "modify its own skills, memory, and operational logic at runtime," and its failure-modes doc reports what that produced on the maintainers' own deployment: a confabulation-mitigation script the agent wrote and never invoked during confabulation, a self-model tool that crashed on wrong file paths, and an analogy patch that returned True instead of conclusions. Three attempts, none worked, and the doc's advice is to "validate externally before relying on them." No gate, no vote, no metric — the drift knob the other instances argue about, set to open, with the outcome written down.

The academic frame

The vendor READMEs above now have a survey behind them. self-evolving-agents-survey (Fang et al., arXiv:2508.07407) defines self-evolving agents as those that "automatically enhance their capabilities through interaction data and environmental feedback, rather than relying on static manually-crafted configurations" — bridging "the static capabilities of foundation models with the continuous adaptability required by lifelong agentic systems." Its useful move is to refuse a single definition: it models the field as a feedback loop over four parts (System Inputs, the Agent System, the Environment, and Optimizers), and notes that different techniques evolve different components. So "self-improvement" is a family scoped by what changes — a config, a skill, a tool set, a policy — not one mechanism. It also treats evaluation, safety, and ethics as central to reliability, which lifts the drift risk below from a hunch to a named research problem.

The bridge to research-wiki

This is compound-engineering (cross-wiki: "each unit of work makes the next easier") moved inside the agent — and it echoes the gbrain (cross-wiki) thesis of an LLM accumulating a compounding personal knowledge base. The open risk is drift/quality: self-authored skills can encode mistakes at scale, so this leans on the same review/eval discipline as the agentic-coding-harness (no published longitudinal measure of whether self-improvement stays net- positive — an open question).

First controlled measurement of the compounding claim (2026-09-03)

harnessdev evaluates six models revising their own agent harnesses from downstream execution feedback, over 2,207 instances with held-out tasks withheld from the loop. Every self-runtime creator improves on the feedback it can see; the gains shrink on unseen tasks, top out at +4.44 points (Opus 4.8), and reverse for three of four lineages once the runtime model is fixed to Gemini 3.1 Pro. "Performance often rises and falls across successive revisions."

That does not refute the mechanism. It sizes it, for the first time here, and the size is small enough that this page's vocabulary needs care: self-improvement in this corpus has meant capability that compounds, and the one controlled study of it finds capability that wobbles and mostly does not travel. Compare compound-engineering-plugin, where SkillFlow found the compounding selective rather than absent — the same shape from a different direction.

The direction can be inverted (added 2026-09-05). environment-evolution keeps the generation-over-generation loop and moves it off the agent entirely: what improves is the environment the agent is trained against, deliberately, because the authors argue that letting the model generate its own curriculum from its own failures is what limits the existing approach. Set beside SkillFlow's selective compounding, this spoke now has two results pointing the same way — the gains come from curation the agent did not do.

A fourth site for the loop (added 2026-09-16). dream-rsi leaves the agent, its skills and its environment alone and improves the exploration policy that tells the agent which candidate to expand next. Feedback comes from replaying candidate policies over the trees earlier runs already built, so the loop pays no new agent calls to score a strategy. The reported wins are mostly the same result in fewer generations — a cost claim, which after harnessdev is the safer kind. scienceide (added {D}) is the environment site again, with real repositories as the source of difficulty; its authors state that the RL gain does not show transfer.

dream-rsi · environment-evolution · hermes-agent · hermes-learn-skill-capture · adk · gstack · agent-skills · durable-agents · compound-engineering · self-evolving-agents-survey · loop-engineering · omega-framework · auto-browser