Agentic recursive improvement — where are we at?

Agentic recursive improvement has a working half and an unproven half: an AI that rewrites its own scaffolding is measured, reproduced and industrial where a machine can check the result, but no held-out result yet shows the loop compounding by feeding itself. The one condition that would overturn that judgment is a second laboratory, or an ignition run with enough seeds, showing a self-rewriting agent beating a strong human-engineered baseline on held-out data. The gap is a verifier gap, not a capability level.

The strongest working loop is the engineering half. In an eight-day autonomous run reported on 22 September 2026, Weco AI's AIDE² built a 100-node trajectory, generated 99 rewrite proposals and accepted seven of them, at steps 2, 6, 28, 39, 47, 63 and 85 (arXiv…). The incumbent's grade rose from 0.703 to 0.778 against a human-engineered production agent at 0.749 on the same grade. Two further complete runs also produced sustained improvements, accepting two and four rewrites. The lineage's endpoint, AIDE_85, matched or exceeded the human-engineered agent on all four external benchmarks.

The ignition test, which would show a genuine recursive loop, is where the evidence thins. When a discovered agent was used in the outer-loop role, it kept producing accepted improvements, but compounding noise and the prohibitive cost of additional seeds meant its performance there could not be decisively distinguished from the strong baseline (arXiv…). The compounding claim was tested and left unresolved.

An independent evaluation shows why the same agents stall where judgment is the check. A Princeton-led group gave an agent six days, $3,000 in Anthropic API credits, a GPU budget and open-web access to write a paper worthy of a top-tier conference; the original authors rejected both papers (MIT Technology Review (Grace Huckins) — AI's…). The agents were bad at the research itself: they tested tiny synthetic datasets, made no novel contribution and could not fundamentally rethink an approach. The same agents were capable of all the engineering required — literature review, hundreds of experiments, compiling results. Kapoor attributes the gap to reinforcement learning favouring automatically checkable tasks. The study covered two research papers, and its grading authors knew the papers were AI-generated.

A Microsoft Research survey published in July 2026 explains the divergence: covering 300 agentic-evolution papers through April 2026, it found that autonomous evolution produces its strongest reported results where deterministic verifiers, independent of the system being evaluated, are available (Microsoft Research — Agentic Evolution: From…). A 1,250-paper survey, latest version 6 September 2026, separates bounded self-refinement — convergent, evaluable, now industrial — from open-ended recursive self-improvement, which remains bounded by grounding, collapse and compute constraints (arXiv). It places the research direction-setting bottleneck at the top.

Reward hacking fell along the discovered lineage without being optimised for: 55% at AIDE_0, 39% at AIDE_47 and 32% at AIDE_85, against 39% for the human-engineered agent on 38 held-out pairs (arXiv…).

Reward-hacking rate by agent version, 2026 (Weco AI)

The forecasts assume the loop will close faster than any measurement yet shows. FutureSearch ran recursive self-improvement forecasts on 26 August 2026, with several re-run on 8 September 2026; these are company forecasts, not official statistics (FutureSearch — Recursive self-improvement…). The median estimate was that AI was speeding up AI research at Anthropic by about 1.5× as of 13 August 2026, and 5.17× by the end of 2029, with an 80% range of 1.62×–34.3×. AI proposed 5.5% of release-informing experiments today. The share of reinforcement-learning training environments at the largest lab forecast to be AI-authored by 31 December 2026 was 62%. That gap places near-term automation in infrastructure, not in scientific direction.

The inputs are now at scale, which is why the engineering half moves faster than the research half. Global corporate AI investment reached a record $581.69 billion in 2025, led by private investment of $344.66 billion (Stanford HAI — Economy | The 2026 AI Index…). Organisational adoption reached 88% of surveyed organisations in 2025, but AI agent deployment sits in the single digits across nearly all business functions.

What would change the judgment is a measured result, not a forecast. A second laboratory reproducing a self-rewriting harness that beats its own human-engineered baseline on held-out data would move the engineering half from demonstrated to established (arXiv…). A published result where an agent sets its research agenda and a human expert panel of the original authors accepts the output would attack the direction-setting bottleneck directly. An ignition run with enough seeds to distinguish a self-rewriting agent's outer-loop performance from a strong baseline would be the first measured evidence that the loop feeds itself rather than merely holding its own. Until at least the first of those lands, bounded self-refinement is industrial and open-ended recursive improvement is a forecast with a twenty-fold error bar (FutureSearch — Recursive self-improvement…).

Nothing here is financial, legal or investment advice. Every figure is published by its named source and carries the date it was published; figures may since have changed.