A Taxonomy of Self-evolving Agents
A practitioner’s blog post that cuts the field of self-evolving agents along a single axis — what evolves — into three loci: model, harness, and artifact. Starting from “Agent = Model + Harness,” it adds artifacts (the agent’s outputs: discovered kernels, generated papers, robot policies) as a third, less-discussed locus, then sorts current work into three levels: artifact iterative optimization, harness self-improvement, and model learning without gold answers. The author is a computer-vision/agents researcher writing informally; two of the cited systems (Eevee, Alita-G) are the author’s own work, flagged as such in the text.
The three levels
Artifact iterative optimization — the agent improves an external output, not itself. Human sets target and evaluation criteria; the agent loops propose → check until criteria are met (AlphaEvolve for algorithm discovery; Analemma AI’s FARS auto-research run — 417 hours, 166 papers, ~$180k; Recursive Superintelligence’s GPU-kernel search). The post’s historical framing: this is neural-architecture-search with the hand-designed operator set and search heuristic both replaced by the LLM, which widens the search space and improves the searcher at once; longer agent horizons (from <5 tool calls in 2024’s LLaVA-Plus to hours of autonomy now) made the loop practical. The frontier is physical environments — robots (NVIDIA), bio labs (LabOS), quantum materials (Qumus).
Harness self-improvement — the agent modifies its own components without touching weights, motivated by training cost. Two sublevels: prompt/memory (GEPA, ACE playbooks, Mem0 — “learning” is a fair label because a harness update is functionally like a parameter update) and tools/skills (Alita, Mem-UI; skills as formalized by Claude Code, framed as context management — reusable procedure without carrying detail in context). Single-agent harnesses stop scaling as playbooks and tool sets grow — retrieval gets noisy and semantics collide — motivating multi-agent self-evolution: per-domain expert agents plus a router (Eevee, Alita-G). Routing is named as the key bottleneck, needing strong base models; the post’s aphorism is that human experts’ most valuable skill is also routing.
Model learning without gold answers — weight updates when only questions, weak signals, or environment access exist. Three sources of signal: pseudo-labels or model-internal confidence (self-training, TTRL, R1-style RL when signals become rewards), self-play and environment interaction grouped as one category since an opponent is just environment (SPIN, Absolute Zero, early-experience learning), and test-time training as an architectural special case (sequence models doing implicit gradient updates at inference; DeltaNet lineage). A section on continual learning notes the older catastrophic-forgetting framing (replay still the main remedy) and observes that terms drift: today’s “continual learning” in LLM discourse often means self-evolving agents, the way “multi-modal” meant captioning in 2017, CLIP in 2021, VLMs in 2023.
Synthesis and closing argument
The boundary is admittedly blurry — optimizing an artifact well may require improving the harness; a saturated harness motivates weight updates (SIA cited as early combined work) — and the post predicts the three loops merge: better models → better harnesses → faster artifact search → new data for model learning. Value should be measured outside the agent system, by real-world artifacts. The distilled takeaway is a three-question test for any self-evolving system: what evolves, what feedback drives it, and where does the loop close — benchmarks, code, science, or the physical world.
Assessment
Category: informal position/organization piece, not a survey — the citation base is a few dozen exemplar systems, chosen for illustration.
- Durable: the three-locus cut and the three questions, both cheap to apply to any concrete system.
- Era-bound: the specific system roster and the “skills are now de facto standard” observations; casual factual claims (FARS run costs, dates) are repeated from primary sources without verification.
- In this library: read against gao2025, the post’s model/harness split is a coarsening of the survey’s four “what evolves” loci (weights vs context+tools+architecture), and its third level maps onto the survey’s population-based/evolutionary paradigm. The genuinely additive moves are (1) elevating the artifact to a first-class evolution locus, which cleanly separates agents-that-improve-outputs from agents-that-improve-themselves — a distinction the survey’s agent-centric formalism handles awkwardly — and (2) the “where does the loop close” question, a sharper deployment-facing probe than the survey’s when/how axes.
Honest about its own blur — the taxonomy is offered as a way to stop arguing about names, not as a formal partition.