Early-Signal Detector

architecture · first seen 5 Oct 2026

self-evolving agents

Also: agent self-evolution, harness self-evolution, retrospective adaptation, coding agent self-evolution

LLM-based agents that autonomously modify their own runtime instructions or tools to adapt to changing conditions.

Self-evolving agents are LLM-based agents that improve from their own experience by rewriting persistent state outside the model weights, such as skill memories, prompts, tool logic or the runtime harness, and sometimes by generating their own training tasks. Recent papers concentrate on making that loop safe and controllable. One paper confines self-modification to the harness so each change in a credit pipeline is a reviewable diff, and another gates memory updates with verifier-based falsification. A security paper shows how distilled skills can be poisoned. Other work applies the idea to medical reasoning, VLM skills and multi-user memory. A few GitHub frameworks such as EvoHarness-RL and EvoSprig package it for developers.

If agents keep adapting at deployment time without retraining, deciding which changes get admitted, audited and protected from poisoning becomes a central engineering and governance problem. Harness-level evolution could make continual agent improvement practical and reviewable in regulated settings.

Analyst note (emerging), 17 Nov 2026.

Mentions per week

01019 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 18 on 16 Nov 2026 · 16 mentions in 28 days · 16 independent groups · code → papers · first seen 6 weeks ago · acceleration +1.45

Earliest sources

Latest sources

Mentioned together with

on-policy self-distillationagent memorySWE-benchexecutable action constraintsworld modelgroup relative policy optimizationagent skillsGEPAbackdoor attackvision-language-action model