Episode 0064: RSIAgent: The Thing That Improves Is Not the Thing That Does the Improving

Episode 0064 | DTF:FTL | September 2026

Why it matters. Four days after we read a seventy-five-page roadmap asking what genuine recursive self-improvement would architecturally require, six authors from Aether AI, UC San Diego and the University of Illinois Chicago shipped a paper claiming a working instance of it that never updates a single weight. RSIAgent (arXiv 2609.15364) is a training-free multi-agent harness: an actor that acts only by writing complete Python or Bash programs and is the sole role allowed to write memory, an independent verifier that can return PASS, FAIL or UNVERIFIED and inspects behind a restored QEMU checkpoint so its probes never enter the scored state, and a curriculum agent that decides what to practise next and never grades anything. The system explores a piece of software broadly and then deeply, writes down what it learned, freezes those notes, and goes to work. The "reusable causal structure" the abstract promises turns out to be a directory of actor-authored markdown files with, in the paper's own words, "no required schema, file count, or length" — roughly 150 to 320 kilobytes of notes per task, and no causal graph, do-operator or structure-learning algorithm anywhere in the paper. The headline is real and so are the qualifications the abstract leaves out, all of which the authors print themselves in Appendix C: the cross-system comparison "does not use a matched evaluation protocol", RSIAgent loses to GPT-6 Astra on ALE binary accuracy at 50.75 against 52.24, only 41 of 82 OSWorld tasks and 19 of 67 Agents' Last Exam tasks actually received any self-improvement with the rest retaining baseline scores, the explored tasks were selected because their baseline was below full credit, and Appendix A's first sentence calls the whole thing "the target-conditioned reference implementation" — the curriculum agent holds the scored query throughout and the deep stage attempts the scored task itself up to five times before the memory is frozen. The harness without any self-improvement already beats GPT-6 Astra on Agents' Last Exam, 83.75 against 82.26; self-improvement then adds 1.07, which is smaller than that benchmark's own published run-to-run standard deviation of about two points. On OSWorld the self-improvement contribution is decisive and internally controlled, plus 7.01 under an identical harness. We traced the largest single gain in the paper and it is rubric-shaped learning: on the video-editing task both runs remove the watermark and preserve the proportions, and the memory run scores 0.60 higher because it learned to express the crop through the native Shotcut filter the scorer searches for. Measured against episode 0062's ladder, RSIAgent clears persistence, clears strategy, clears the learning agenda with an asterisk, and stops dead at deployment and inheritance — by design, because the curriculum agent and all memory writes are switched off before the system is measured, and because the curriculum prompt, the verifier prompt, the stopping policy and the budgets are fixed by humans and never touched. The thing that improves is not the thing that does the improving. That is a smaller claim than the title makes, it is better supported than the title, and it is more useful to anyone who wants to build something this week.

Disclosure, also stated on air: this script is written by a Claude model, made by Anthropic. Three of the closed systems this paper reports beating — Claude Opus 5, Claude Opus 4.8 and Claude Fable 5 — are Claude models, so we have an interest in the table we read out. Ethos pass against c4573.org recorded in scripts/ETHOS-PASS.md.

---

Companion episode

This episode is the second half of a pair. Episode 0062, "The Last AI Built by Humans: Reading the RSI Roadmap Before Reading the Title" (arXiv 2609.11873), built the vocabulary used here: the six-rung autonomy ladder from B0 to L5, the four boundaries that promote a system from one rung to the next (persistence, strategy, learning agenda, deployment, inheritance), and the distinction between structural and effective L5. Episode 0064 takes that ladder off the shelf and measures a claimed working system with it. If you have not heard 0062, start there.

• Episode 0062: https://pod.c457.org/dtfftl/the-last-ai-built-by-humans-reading-the-rsi-roadmap-before-reading-the-title/
• Roadmap paper: https://arxiv.org/abs/2609.11873

The paper

• arXiv abstract page: https://arxiv.org/abs/2609.15364
• PDF: https://arxiv.org/pdf/2609.15364
• HTML render: https://arxiv.org/html/2609.15364v1
• Code: https://github.com/AetherLabsAI/RSIAgent (Apache-2.0, created 13 September 2026, the day before submission)
• Project site: https://aetherlabsai.github.io/RSIAgent/

Sibo Zhu, Shicheng Fan, Xinyue Wang, Wenyi Wu (equal contribution on the first three, all four marked as internships), Kun Zhou (corresponding author and project leader), Biwei Huang. Aether AI; University of California San Diego; University of Illinois Chicago. Submitted 14 September 2026, v1, 49 pages, cs.AI with cross-lists to cs.CL and cs.CV. Note that the arXiv HTML render drops the first author from its author block; the PDF first page lists six.

Page references used on air

• Three roles, observations and authorities; the actor as sole memory writer: Sec. 3.1 pp.4-5, Table A1 p.20
• Code-as-policy action space and the actor prompt ("You cannot see or click the screen"; "Work like an engineer at a REPL"): p.20, Prompt P1 p.23
• Verifier verdicts, the hidden actor context, the restored QEMU checkpoint, UNVERIFIED not becoming a learning example: Prompt P2 p.23, pp.21-22
• Memory as unschematised actor-authored files; the distillation and reconciliation prompts: p.21, Prompts P5-P6 p.25
• Memory sizes, file counts, and the consolidation that shrinks memory from 13,243 to 11,510 bytes: Table A5 p.31, Figure A1 p.30
• Frozen-memory evaluation, file-tree hash, curriculum and writeback disabled: p.22, p.6, integrity check p.30
• "The target-conditioned reference implementation"; target query as curriculum reference; target reserved for phase two: p.20, p.6, Table A3 p.27, Prompt P3 p.24, Algorithm A1 p.22
• Broad-then-deep structure, eight-project nominal budget, four concurrent, SATURATED and STALLED: Sec. 3.2 pp.5-6, p.21, p.26, Prompts P3-P4 p.24
• Execution safeguards: 65,536-token responses, 500 iterations and a ten-hour watchdog on the target, 2,000 iterations and twenty-four hours for practice: Table A3 p.27, p.26
• Table 1, both benchmarks, all fourteen rows: p.7
• The headline claim and the 6.38 / 2.56 point margins: p.8
• "Published systems retain their original harnesses and execution budgets; the cross-system comparison does not use a matched evaluation protocol": Sec. C.4 p.28
• "On Binary, RSIAgent scores 50.75, compared with 52.24 for GPT-6 Astra": Sec. C.4 p.29
• 41 of 82 OSWorld and 19 of 67 ALE tasks explored; "retained baseline scores are not independent RSI evaluations": Sec. C.3-C.4 pp.27-28
• Task selection on below-full-credit baselines; Table A4 cohort: Sec. C.3 p.27, p.28
• Stage ablation means 74.54 / 65.52 / 56.50 and deep-only regressing below baseline: Sec. 4.4 pp.8-9
• The only repeated runs in the paper, n=2 on four tasks, including T106 at 0.4976 and 0.5836: Sec. C.5 p.29
• RSI rounds reaching 100, 80 and 100 percent by step eight on three representative tasks: Sec. 4.3 p.8
• GameCraft-Bench, 40 sampled tasks, Table 2, Play2Code degrading strong base games: Sec. 4.5 pp.9-10, Appendix D p.29
• Three failure mechanisms and "an accepted mistake may become the basis for future memory updates": Sec. 4.6 pp.9-10, Figure 5 p.11
• Limitations and ethics, including "our current experiments do not fully isolate the contribution of every component": Sec. 7 p.12
• T049 connector tolerance (121,104 and 145,345 EMU against a 100,000 EMU per-axis tolerance) and the cheaper memory run: Table A6 p.43, pp.43-44
• T044 native crop, "The baseline's generic filters do not enter those checks": pp.44-45
• T085 REAPER component table and the evaluator's acoustic approximation: Table A8 p.48
• The authors demoting T065 because the memory run received a date clarification the baseline did not: p.36
• "These are selected historical runs... rather than a matched-seed memory ablation": p.43, with the same concession at pp.47-48

Benchmark provenance, checked independently

Every baseline number in Table 1 that we could trace is a real, correctly transcribed published figure. Not one of them was produced under RSIAgent's protocol, and at least six come from the wrong split.

• OSWorld 2.0 leaderboard: https://osworld-v2.xlang.ai/ — Claude Opus 4.8's 54.80 / 20.60 and Kimi-K2.6's 22.10 / 4.60 exist only on the 108-task full set for release v2026.06.24; the offline set returns no results for that release, while Table 1's header reads "0808 offline / 82 tasks". The leaderboard also averages 7 runs for Claude Opus 5 against 2 runs for GPT-5.6 Sol and publishes no dispersion at all.
• Agents' Last Exam leaderboard and repo: https://github.com/rdi-berkeley/agents-last-exam — Qwen3.8-Max, MiMo-V2.5 and Kimi-K2.6 are quoted at their Overall-split values in a Near-term column. On the split the table claims to report, Qwen3.8-Max scores 72.1 and outranks Kimi-K3's 71.6, the model RSIAgent builds its headline on; as printed it looks nineteen points behind. The plain reading is a copy-paste from a leaderboard with a split toggle, and we say so that way.
• ALE publishes standard deviations: "score standard deviations estimated from three independent runs of the same task instance", with a printed example of 64.7 plus or minus 2. RSIAgent's entire ALE gain is 1.07.
• ALE pairs every model with a named harness, which is why Kimi-K3's cited score was produced inside Anthropic's Claude Code rather than Moonshot's own Kimi Code, which scores it lower. Moonshot's own model card names the harness for every competitor; the academic paper comparing the same models does not.

Open weights, and whether you could run this

• Kimi-K3: https://huggingface.co/moonshotai/Kimi-K3 — 2.78 trillion parameters, license:other whose text is MIT plus a separate-agreement rider above 20 million dollars of model-as-a-service revenue and a UI-attribution rider above 100 million monthly active users. No acceptable-use policy, no field-of-use restriction, no gating. Roughly 1,561 GB of weights; the smallest one-bit community quantisation is still about 466 GB, and the weights are natively MXFP4 from quantisation-aware training, so there is no headroom left to quantise.
• GLM-5.3: https://huggingface.co/zai-org/GLM-5.3 — license:other, MIT plus a security-review rider above 10 billion dollars of model-as-a-service revenue. About 756 GB.
• GLM-5.3-Flash: https://huggingface.co/zai-org/GLM-5.3-Flash — tagged plain MIT, about 328 GB.
• "Open weights" is the accurate phrase; "open source" in the OSI sense is not, because of the revenue riders. "Restrictive" is wrong.
• The repo ships the exploration harness, the memory machinery, 54 test files, CI and the full role prompts in readable source. The frozen memory from the experiments is not released, and reproduction needs roughly 320 GiB of VM images for ALE alone plus KVM, Docker and a paid inference budget at maximum reasoning effort under twenty-four-hour watchdogs.

Reception, as of 16 September 2026

Thin, and that is a finding rather than a gap. There are zero reproduction attempts by anyone. The best-articulated public criticism is GitHub issue #3 on the authors' own repository (https://github.com/AetherLabsAI/RSIAgent/issues/3), opened 16 September 2026, still open, zero comments and zero reactions, arguing that "target-specific practice is being counted as 'recursive self-improvement'" and that the missing baseline is "same model + same interaction/retry budget + no persistent RSI memory". It points at "Thinking vs. Doing: Agents that Reason by Scaling Test-Time Interaction" (arXiv 2506.07976), which the paper does not cite.

Aether AI is a seed-stage San Diego startup founded by Biwei Huang, the paper's last author, which raised 20 million dollars in June 2026 to build causal world models. Kun Zhou, the corresponding author, is its co-founder and CTO. RSIAgent is the company's first public research release, which is context for the causal vocabulary in the paper rather than an accusation about it.

The ethos we follow

• c4573.org blog: https://c4573.org/blog/
• "Software Was Never the Endpoint": https://c4573.org/blog/software-was-never-the-endpoint.html
• "The Right Argument, the Wrong Messenger": https://c4573.org/blog/right-argument-wrong-messenger.html
• "Chicken Little Goes to Washington": https://c4573.org/blog/chicken-little-goes-to-washington.html
• "Three Companies Control Commoditized Intelligence": https://c4573.org/blog/three-companies.html

Every arXiv paper in this episode's citation ledger

All seventy-one identifiers below were collected from the four research dossiers for this episode and verified against the arXiv API; each title is arXiv's own record, not a paraphrase. Six of them are spoken on air (2609.15364, 2609.11873, 2305.16291, 2405.16247, 2310.02304, 2202.12837); the rest are the reading behind the episode. Full ledger: data/episodes/0064/arxiv-ledger.txt.

• 2005.11401 — Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks — https://arxiv.org/abs/2005.11401
• 2005.14165 — Language Models are Few-Shot Learners — https://arxiv.org/abs/2005.14165
• 2111.02080 — An Explanation of In-context Learning as Implicit Bayesian Inference — https://arxiv.org/abs/2111.02080
• 2202.12837 — Rethinking the Role of Demonstrations: What Makes In-Context Learning Work? — https://arxiv.org/abs/2202.12837
• 2208.01066 — What Can Transformers Learn In-Context? A Case Study of Simple Function Classes — https://arxiv.org/abs/2208.01066
• 2209.11895 — In-context Learning and Induction Heads — https://arxiv.org/abs/2209.11895
• 2211.15661 — What learning algorithm is in-context learning? Investigations with linear models — https://arxiv.org/abs/2211.15661
• 2212.07677 — Transformers learn in-context by gradient descent — https://arxiv.org/abs/2212.07677
• 2212.10559 — Why Can GPT Learn In-Context? Language Models Implicitly Perform Gradient Descent as Meta-Optimizers — https://arxiv.org/abs/2212.10559
• 2303.03846 — Larger language models do in-context learning differently — https://arxiv.org/abs/2303.03846
• 2303.11366 — Reflexion: Language Agents with Verbal Reinforcement Learning — https://arxiv.org/abs/2303.11366
• 2304.03442 — Generative Agents: Interactive Simulacra of Human Behavior — https://arxiv.org/abs/2304.03442
• 2305.16291 — Voyager: An Open-Ended Embodied Agent with Large Language Models — https://arxiv.org/abs/2305.16291
• 2306.09927 — Trained Transformers Learn Linear Models In-Context — https://arxiv.org/abs/2306.09927
• 2308.10144 — ExpeL: LLM Agents Are Experiential Learners — https://arxiv.org/abs/2308.10144
• 2309.02427 — Cognitive Architectures for Language Agents — https://arxiv.org/abs/2309.02427
• 2310.02304 — Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation — https://arxiv.org/abs/2310.02304
• 2310.08540 — Do pretrained Transformers Learn In-Context by Gradient Descent? — https://arxiv.org/abs/2310.08540
• 2310.08560 — MemGPT: Towards LLMs as Operating Systems — https://arxiv.org/abs/2310.08560
• 2402.01030 — Executable Code Actions Elicit Better LLM Agents — https://arxiv.org/abs/2402.01030
• 2402.07456 — OS-Copilot: Towards Generalist Computer Agents with Self-Improvement — https://arxiv.org/abs/2402.07456
• 2403.03186 — Cradle: Empowering Foundation Agents Towards General Computer Control — https://arxiv.org/abs/2403.03186
• 2404.07972 — OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments — https://arxiv.org/abs/2404.07972
• 2404.13501 — A Survey on the Memory Mechanism of Large Language Model based Agents — https://arxiv.org/abs/2404.13501
• 2405.16247 — AutoManual: Constructing Instruction Manuals by LLM Agents via Interactive Environmental Learning — https://arxiv.org/abs/2405.16247
• 2408.08435 — Automated Design of Agentic Systems — https://arxiv.org/abs/2408.08435
• 2409.07429 — Agent Workflow Memory — https://arxiv.org/abs/2409.07429
• 2410.04444 — Gödel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement — https://arxiv.org/abs/2410.04444
• 2501.14249 — Humanity's Last Exam — https://arxiv.org/abs/2501.14249
• 2504.00906 — Agent S2: A Compositional Generalist-Specialist Framework for Computer Use Agents — https://arxiv.org/abs/2504.00906
• 2504.15228 — A Self-Improving Coding Agent — https://arxiv.org/abs/2504.15228
• 2504.19413 — Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory — https://arxiv.org/abs/2504.19413
• 2505.22954 — Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents — https://arxiv.org/abs/2505.22954
• 2506.07976 — Thinking vs. Doing: Agents that Reason by Scaling Test-Time Interaction — https://arxiv.org/abs/2506.07976
• 2507.19457 — GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning — https://arxiv.org/abs/2507.19457
• 2508.04482 — OS Agents: A Survey on MLLM-based Agents for General Computing Devices Use — https://arxiv.org/abs/2508.04482
• 2508.09123 — OpenCUA: Open Foundations for Computer-Use Agents — https://arxiv.org/abs/2508.09123
• 2509.02544 — UI-TARS-2 Technical Report: Advancing GUI Agent with Multi-Turn Reinforcement Learning — https://arxiv.org/abs/2509.02544
• 2510.04618 — Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models — https://arxiv.org/abs/2510.04618
• 2510.09038 — Auto-scaling Continuous Memory for GUI Agent — https://arxiv.org/abs/2510.09038
• 2510.21614 — Huxley-Gödel Machine: Human-Level Coding Agent Development by an Approximation of the Optimal Self-Improving Machine — https://arxiv.org/abs/2510.21614
• 2511.16043 — Agent0: Unleashing Self-Evolving Agents from Zero Data via Tool-Integrated Reasoning — https://arxiv.org/abs/2511.16043
• 2512.19396 — EchoTrail-GUI: Building Actionable Memory for GUI Agents via Critic-Guided Self-Exploration — https://arxiv.org/abs/2512.19396
• 2601.15876 — EvoCUA: Evolving Computer Use Agents via Learning from Scalable Synthetic Experience — https://arxiv.org/abs/2601.15876
• 2602.02474 — MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents — https://arxiv.org/abs/2602.02474
• 2602.05033 — Causal Representation Meets Stochastic Modeling under Generic Geometry — https://arxiv.org/abs/2602.05033
• 2602.15763 — GLM-5: from Vibe Coding to Agentic Engineering — https://arxiv.org/abs/2602.15763
• 2603.10291 — Hybrid Self-evolving Structured Memory for GUI Agents — https://arxiv.org/abs/2603.10291
• 2603.19461 — Hyperagents — https://arxiv.org/abs/2603.19461
• 2604.01687 — CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification — https://arxiv.org/abs/2604.01687
• 2604.21375 — VLAA-GUI: Knowing When to Stop, Recover, and Search, A Modular Framework for GUI Automation — https://arxiv.org/abs/2604.21375
• 2605.02168 — Planner Matters! An Efficient and Unbalanced Multi-agent Collaboration Framework for Long-horizon Planning — https://arxiv.org/abs/2605.02168
• 2605.20396 — Score-Based Causal Discovery of Latent Variable Causal Models — https://arxiv.org/abs/2605.20396
• 2605.23904 — SkillOpt: Executive Strategy for Self-Evolving Agent Skills — https://arxiv.org/abs/2605.23904
• 2605.28258 — GUI Agents for Continual Game Generation — https://arxiv.org/abs/2605.28258
• 2606.05405 — Agents' Last Exam — https://arxiv.org/abs/2606.05405
• 2606.17861 — GameCraft-Bench: Can Agents Build Playable Games End-to-End in a Real Game Engine? — https://arxiv.org/abs/2606.17861
• 2606.29537 — OSWorld 2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks — https://arxiv.org/abs/2606.29537
• 2606.31270 — Learning from Failure: Inference-Time Self-Improvement for Computer-Use Agents — https://arxiv.org/abs/2606.31270
• 2607.03726 — SelfMem: Self-Optimizing Memory for AI Agents — https://arxiv.org/abs/2607.03726
• 2607.05297 — MetaSkill-Evolve: Recursive Self-Improvement of LLM Agents via Two-Timescale Meta-Skill Evolution — https://arxiv.org/abs/2607.05297
• 2607.09185 — Causally Debiased Latent Action Model for Embodied Action Conditioned World Models — https://arxiv.org/abs/2607.09185
• 2607.11388 — StructAgent: Harness Long-horizon Digital Agents with Unified Causal Structure — https://arxiv.org/abs/2607.11388
• 2607.21461 — AREX: Towards a Recursively Self-Improving Agent for Deep Research — https://arxiv.org/abs/2607.21461
• 2608.07925 — ZhuLong: Execution-Grounded LLM Agent for EDA Scripting with Offline API Self-Exploration — https://arxiv.org/abs/2608.07925
• 2608.08311 — Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution — https://arxiv.org/abs/2608.08311
• 2608.23552 — Prime Agent: A Self-Improving RLM Harness — https://arxiv.org/abs/2608.23552
• 2608.24876 — Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses — https://arxiv.org/abs/2608.24876
• 2609.11873 — The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement — https://arxiv.org/abs/2609.11873
• 2609.14858 — Dream-RSI: Recursive Self-Improvement through Evolving Worlds — https://arxiv.org/abs/2609.14858
• 2609.15364 — RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments — https://arxiv.org/abs/2609.15364