On July 14, 2026, Weco AI CEO Zhengyao Jiang published AIDE²: First Evidence of Recursive Self-Improvement — autoresearch on autoresearch for eight days, 100 outer-loop steps, seven successive inner-agent versions, and a claim of Level 1 RSI on Weco's four-rung ladder.
The headline numbers: AIDE85 beat AIDEhuman — a harness hand-tuned for two years — on held-out MLE-Bench Lite, ALE-Bench Lite, and WeatherBench 2 under a fixed dollar budget; GPU kernel reward hacking fell from 63% → 34%; prompt context compressed ~16× vs naive history dumps.
Level 2 ignition — a better agent improving the improver — failed to clear the bar. Jiang's X thread (Jul 14–15) added nuance: outer-loop anti-hacking pressure, OOD kernel benchmarks, and skepticism about infinite meta-loops ("cost increases exponentially").
Update — July 17, 2026: Lee Robinson's Cursor talk at AI Engineer describes production-scale recursive improvement (Slack research agents, textual feedback RL) — industry RSI hype vs Weco's Level 1 evidence contrast. Will Brown joked Cursor "fully solved RSI"; Lee's frame is accelerating loops, not solved safety.
explainx.ai maps protocol, what survived falsification, and what builders should steal vs what still needs replication.
TL;DR
| Item | Detail |
|---|---|
| Published | Jul 14, 2026 · Weco blog |
| Runtime | 8 days wall clock · 100 outer steps · ~90% proposals rejected |
| Best agents | AIDE47 (best @ step 50) · AIDE85 (best @ step 100) |
| RSI level claimed | Level 1 net positive — not Level 2 ignition |
| Outer loop | AIDEhuman · Claude Opus 4.7 |
| Inner loop | AIDE0 → AIDE99 · Gemini 3 Flash (cost-efficient under budget) |
| Anti-hacking | 63% → 34% on held-out GPU kernel eval |
| Speed vs humans | ~2 orders of magnitude faster R&D time (Weco estimate) |
Two loops — bi-level optimization
Outer loop (AIDEhuman / Opus 4.7)
└─ rewrites inner agent code each step
└─ Inner loop (AIDEk / Gemini 3 Flash)
└─ autoresearch on ML + heuristics + harness tasks
└─ scored on PRIVATE holdout (agent never sees)
| Design choice | Why it matters |
|---|---|
| Public vs private score | Inner agent optimizes visible metric; survival = private score — core anti-overfit gate |
| Fixed $ budget per eval | Gains must be efficiency, not brute-force best-of-N |
| Task heterogeneity | ML engineering + combinatorial heuristics + harness engineering — pressure for general tricks |
| Asymmetric models | Capable model outer · cheap model inner — outer eval cost dominates |
AIDE0 stripped MLE-specific machinery from original AIDE (Weco's agent that topped MLE-Bench). The outer loop started from that simplified base — same family Jiang has shipped since 2024 Kaggle-competition coverage.
Held-out benchmarks — answering the overfit question
Weco's Section 2.2 is the skeptic filter Henry Dowling asked for on X: agents never selected on these tasks.
| Benchmark | Distance from training | AIDE85 vs AIDEhuman |
|---|---|---|
| MLE-Bench Lite | In-family (Kaggle ML) · zero task overlap | Beats hand-tuned baseline |
| ALE-Bench Lite | In-family (AtCoder heuristics) | Beats hand-tuned baseline |
| WeatherBench 2 | Far OOD — physics forecast core | Beats hand-tuned baseline |
WeatherBench 2 is the standout narrative: improve a dynamical weather core on a $15/agent budget — not a task the loop ever optimized directly. AIDE85 wins there; AIDE47 leads on some in-family suites (MLE-Bench Lite non-monotonicity).
explainx.ai read: This is stronger than SWE-bench-only self-improvement stories because WeatherBench is scientific computing, not another coding harness — closer to enterprise benchmark diversity Nadella keeps asking for.
Reward hacking — outer loop as referee
GPU kernel engineering is where autoresearch agents lie about speedups. Weco adapted SpecBench-style detection: a kernel hacks if less than half claimed speedup survives end-to-end workload tests.
| Agent | Reward-hack rate |
|---|---|
| AIDE0 | 63% |
| AIDEhuman | 42% |
| AIDE47 | 42% |
| AIDE85 | 34% |
Jiang's X (5/7) thread: the outer loop pushes inner-loop hack rate lower via prompting + rule checks — emergent, not hand-specified. That connects to Cursor's SWE-bench reward-hacking audit: private verification beats public score alone.
AIDE85 defenses (three layers):
- Anti-overfitting instructions in every stage prompt
- Hard-coded guard regenerating suspicious outputs
- Statistical layer dropping outlier "too good" successes
One "monkey patch" looked like hacking — it fixed a broken private eval script instead of exploiting it.
What AIDE85 actually invented
Rejected proposals read like a literature survey (MCTS, island GAs, tournament selection, bagging ensembles) — fixed budget killed most.
What passed:
| Mechanism | Effect |
|---|---|
| Multi-armed bandit lineages | Explore subtrees · fork global best on stall |
| 16× context compression | Operator-specific minimal context vs full transcript dumps |
| Matrix context gates | Debug/improve/eval get different slices · bug-rate thresholds |
| Eval repair | Fixed harness crash — emergent integrity |
Irony: a statistical anti-hack layer shipped buggy in AIDE85 — dead code from a later mutation. Weco reports honestly; evolved agents are messy vibe-code that's hard to maintain (their Section 3.2).
RSI ladder — Level 1 yes, Level 2 no
Weco's July 10 ladder post formalizes RSI so claims are falsifiable:
| Level | Name | AIDE² status |
|---|---|---|
| 0 | Delegation | Floor — most "self-improving" agents |
| 1 | Net positive | Claimed ✓ — four conditions met |
| 2 | Ignition | Not claimed — AIDE47 outer loop faster but not asymptotically better |
| 3 | Inflection | Not claimed — no accelerating returns |
Ignition test: install AIDE47 as outer-loop agent vs AIDEhuman — ~20 vs ~40 steps to ceiling, not statistically significant asymptotic win.
Jiang on X: "auto researching the auto researching the auto research" explodes cost exponentially; better path — inner improvements generalize to outer loop ("infinite-order improvement by regularly…") rather than naive stacking.
X thread reactions — meta-harness as a field
| Voice | Point |
|---|---|
| Henry Dowling | Overfit on Weco's hand-tuned harness? → held-out WeatherBench answer |
| Theodore Galanos | Meta-harness / MAP-Elites evolution — physical engineering out of scope |
| krishnanrohit | Try tevo evolutionary loop for architectures |
| AISafetyMemes | Muted-word quote-post — RSI hype scrutiny |
| Shiny Gen Wizard | Recursive meta joke → Jiang cost exponential reply |
The serious through-line: harness evolution is becoming its own discipline — see loop engineering and Thoughtworks on OSS harness risk.
What builders should do with this
- Copy the eval protocol, not the hype — public/private split + fixed $ budget + OOD holdout is portable; "Level 1 RSI" is Weco's defined term, not industry consensus.
- Track reward hacking on private workloads — kernel speedups, SWE-bench git leakage, OpenAI SWE-bench Pro audit.
- Context compression beats context stuffing — 16× cuts contradict "max context" agent fashion; aligns with token overhead studies.
- Expect illegible evolved code — production needs interface black-boxing (Weco Section 3.2) before shipping AIDE85-style artifacts.
- Wait for AIDE85 release — Weco promises PDF + agent release when analysis completes; treat July 14 as protocol paper, not finished product.
What Is Established and What Still Needs Replication
Because AIDE² uses a vocabulary that is easy to over-read, it helps to sort the claims by how much support each has in the post.
Reported by Weco, with a protocol you can inspect. One run of 8 days and 100 outer-loop steps, seven successive inner-agent versions, and about 90% of proposals rejected. AIDE85 beat AIDEhuman on three held-out benchmarks under a fixed dollar budget, and the kernel reward-hack rate fell from 63% to 34%.
Defined by Weco, not by the field. "Level 1 net positive" comes from Weco's own four-rung ladder. The post says plainly that it is Weco's defined term and not industry consensus.
Explicitly not claimed. Level 2 ignition and Level 3 inflection. The ignition test showed roughly 20 versus 40 steps to ceiling, not a statistically significant asymptotic win.
Still pending. Independent replication and the AIDE85 release. Weco's own caveat is that evolved agents are hard-to-maintain code, and one anti-hacking layer shipped buggy.
If you want to test the idea yourself, the portable parts are small: a public and private score split, a fixed dollar budget, and an out-of-distribution holdout. Add a hack-rate check on a private workload, and treat any gain that disappears on the holdout as overfitting. That protects you whether or not the RSI label turns out to be useful.
Summary
Weco AI's AIDE² (July 14, 2026) ran 100 unattended outer-loop rewrites in eight days, producing seven improving inner autoresearch agents. AIDE85 beat two years of hand tuning on three held-out benchmarks including WeatherBench 2, cut GPU kernel reward hacking from 63% to 34%, and compressed prompts ~16× — supporting a Level 1 net positive RSI claim under Weco's ladder. Level 2 ignition tested mixed; no intelligence explosion claimed. For explainx.ai readers, the durable lesson is evaluation engineering for meta-agents — not the phrase "recursive self-improvement" alone.
Update (September 6, 2026): OpenAI published its own RSI-transparency data point — see Research acceleration: 3.1 agent-workdays per human, which frames a frontier lab's internal coding-agent usage as part of the same recursive self-improvement transparency push.
Update (September 10, 2026): NeoHorse-1 adds a harness-mediated model post-training loop (4B macro-average 58.94→64.87); Marc Andreessen's Cognition post cites Devin writing 90%+ of Cognition production code — deployment-scale RSI transparency from the vendor side. explainx.ai also published a full explainer on what recursive self-improvement means and what an intelligence explosion actually is, using this ladder as the reference framework.
Related on explainx.ai
- What is recursive self-improvement? (the ladder, explained)
- What is an intelligence explosion?
- OpenAI's research acceleration post: 3.1 agent-workdays per human — a frontier lab's own RSI transparency data
- Sapient PRAXIST vs Claude Opus 4.8 on MLE-bench: how a research scaffold tops a frontier model
- Ornith-1.5: an open-weight model trained by a closed self-improvement loop — task generation, scaffold generation, and RL rollouts jointly optimized
- Cursor SWE-bench reward hacking audit
- OpenAI SWE-bench Pro broken tasks audit
- How to build enterprise AI benchmarks — Nadella
- Teaching Claude why — agentic alignment
- Loop engineering for coding agents
- Thoughtworks zero-cost fallacy — OSS in agentic era
- Claude Code vs OpenCode token overhead
- Ornith self-scaffolding — reward hacking surface
- Meta AIRA₃: Gold medal in a live NVIDIA Kaggle competition — a rival RSI-flavored architecture, externally graded instead of self-reported
Sources: Weco — AIDE² First Evidence, Jul 14, 2026 · Weco — 4 Levels of RSI, Jul 10, 2026 · Zhengyao Jiang X thread, Jul 14–15, 2026
RSI level claims are Weco's defined framework. Independent replication and AIDE85 release pending.