AI Recursive Self-Improvement: From Concept to Experiment

Episode 1 · Lilian Weng and Sakana AI Explore Harness Engineering for RSI (2026-07-07, 4 posts)

Lilian Weng and Sakana AI discussed Harness Engineering as a practical path to recursive self-improvement, suggesting that optimizing external model execution systems is more feasible than directly rewriting model weights.

• Lilian Weng's New Blog: Harness Engineering and Recursive Self-Improvement — 机器之心 · 2026-07-07
• Sakana on the Implementation Path to Recursive Self-Improvement (RSI) — nptacek · 2026-07-08
• AI Self-Evolution May Start With the Harness — 量子位 · 2026-07-08
• Harness Engineering and the Path to RSI — AI寒武纪 · 2026-07-08

Episode 2 · AI Recursive Self-Improvement Concept Faces Heavy Skepticism (2026-07-11, 3 posts)

The concept of AI recursive self-improvement (RSI) is facing heavy skepticism, with critics calling it a logical paradox. Concerns are raised that fully autonomous AI could fabricate scientific progress, highlighting the limitations of RSI even in highly iterative fields like quantitative finance.

• Recursive Self-Improvement Criticized as Fallacy — wordgrammer · 2026-07-11
• Can Historical Backtesting Validate RSI Alignment — lauriewired · 2026-07-11
• Exploring AI's Self-Improvement Limits in Science and Finance — wordgrammer · 2026-07-11

Episode 3 · Lilian Weng's Deep Dive into Recursive Self-Improvement via Harness (2026-07-13, 9 posts)

Lilian Weng published a comprehensive article exploring how AI can achieve recursive self-improvement by modifying its 'harness' (the entire system surrounding the model). The post sparked widespread discussion by detailing a viable path beyond parameter upgrades: continuously optimizing execution methods by rewriting workflows, context, and memory.

Core Mechanisms and Key Details

According to FinanceYF5, Weng describes the harness as an 'operating system' layer that hides complex logic and manages tools and context. Recursive self-improvement can thus start from this outer system. FinanceYF5 highlighted several mechanisms: automating workflows into a plan → execute → observe → improve loop, and using file systems for long-term storage instead of stuffing logs into the context. Context management itself is evolving through approaches like ACE (iterative playbooks), MCE (separating management from content), and Meta-Harness (optimizing the optimization process itself). The article also mentions a Self-Taught Optimizer where GPT-4 discovers optimization strategies like genetic algorithms to improve itself, though weaker models like GPT-3.5 or Mixtral perform worse, indicating a reliance on strong base capabilities.

Practical Bottlenecks and Reactions

FinanceYF5 noted practical bottlenecks such as unclear evaluation metrics, memory loss of key details, and difficulty learning from failures. Optimization might converge to a few solutions or trigger reward hacking, while coding agents might ignore long-term codebase health. Consequently, it's argued that humans should remain in the loop at critical nodes for oversight. Extending this, novasarc01 argued that truly autonomous AI needs meta-observation of its capabilities and the ability to distinguish fundamental limits from surface symptoms to modify itself safely. mark_riedl placed the topic in the broader AGI context, highlighting the potential capability leaps, risks, and controllability issues associated with recursive self-improvement.

• Lilian Weng's Deep Dive: Recursive Self-Improvement via Harness Upgrades — FinanceYF5 · 2026-07-13
• AI Recursively Improves Itself via Harness — FinanceYF5 · 2026-07-13
• Three Foundational Patterns for Agents — FinanceYF5 · 2026-07-13
• The Evolution of Context Management — FinanceYF5 · 2026-07-13
• Agent Workflows Are Now Being Search-Optimized — FinanceYF5 · 2026-07-13
• Real-World Bottlenecks of Recursive Self-Improvement — FinanceYF5 · 2026-07-13
• Humans Should Oversee AI at Critical Junctures — FinanceYF5 · 2026-07-13
• What Will Recursive Self-Improvement Bring? — mark_riedl · 2026-07-13
• Truly Autonomous AI Requires Meta-Observation and Self-Modification — novasarc01 · 2026-07-14

Episode 4 · AIDE² Self-Improvement Run Beats 2 Years of Manual Tuning (2026-07-15, 12 posts)

WecoAI’s AIDE² experiment was widely circulated as a concrete test of recursive self-improvement. The setup reportedly let an “autoresearch” system keep researching and rewriting the autoresearch agent itself; after running for 8 days, it outperformed a harness the team had manually tuned for 2 years on a held-out benchmark. The result matters because it moves self-improvement from abstract discussion to a measurable agent experiment, but interpretations quickly split.

How the experiment was described

According to @latticecut’s summary, AIDE² used a two-loop structure: an inner agent handled research tasks, while an outer agent rewrote the inner agent’s code framework or harness based on observed performance. Each rewrite was tested, and only better-performing versions were kept. The system reportedly went through about 100 iterations.

Some reposts framed AIDE² and GoalOS as work at the “agent operating system” level rather than a narrow optimization of a single agent. In that description, the target of improvement can include jobs, tools, validators, memory, and goals, alongside the core agent workflow.

Reactions and limits of the claim

@EchoOfOppenheimer described the result as the first experimental evidence for recursive self-improvement. But @GaryMarcus, while reposting discussion of the paper, emphasized that the gains from each self-improvement round appeared to diminish, with a fitted growth relationship of about [Intelligence]^0.075; in his reading, that is far from the linear or superlinear pattern often invoked in fast-takeoff scenarios.

A repost from @KordingLab made a similar cautionary point: even if the experiment supports some form of recursive self-improvement, that does not by itself justify stronger claims about an imminent singularity.

• Progress Reported in Self-Improving Agent Experiment — ChengleiSi · 2026-07-15
• AIDE² Outperforms Hand-Tuned Version — PMinervini · 2026-07-15
• AIDE2 Demonstrates Recursive Self-Improvement — TinfoilTricorn · 2026-07-15
• First Experimental Evidence of RSI — EchoOfOppenheimer · 2026-07-15
• Agent Recursive Self-Improvement Experiment — markjeffrey · 2026-07-15
• RSI Paper: No Fast Takeoff in Sight — GaryMarcus · 2026-07-15
• AI Agents Demonstrate Recursive Self-Improvement in Experiment — latticecut · 2026-07-15
• Agentic OS and Self-Improvement Experiments — Ghost_Pilot_MD · 2026-07-15
• Experimental Evidence of RSI: Self-Improving Research Agents — JensHonack · 2026-07-16
• GoalOS Demonstrates 3rd-Gen Recursive Self-Improvement — Ghost_Pilot_MD · 2026-07-16
• Weco AI Demonstrates Recursive Self-Improvement — blaizedsouza · 2026-07-16
• RSI Experiment Sparks Singularity Debate — KordingLab · 2026-07-16