"Recursive Self-Improvement" — Anthropic Institute

Why this is in the vault

Independent, first-party, this-week framing that (1) names the execution-vs-judgment line RDCO's /improve loop already sits behind, and (2) validates today's supervisor-agent / /verify-write-action design as the correct guardrail for a self-editing agent. Shared by founder via iMessage 2026-06-04 ~21:50 ET. (Note: this specific article was NOT previously filed; the April 2026-04-04-recursive-self-improvement-marketing.md is a different, marketing-applied piece.)

The core argument

• RSI defined as "an AI system capable of fully autonomously designing and developing its own successor." Not a containment paper — a "we're closer than you think; build the governance infra now" position (METR/Epoch-style extrapolation, AI-2027-shaped).
• The missing ingredient is judgment/taste in choosing research goals, not execution. Execution is largely solved.
• Evidence (Anthropic self-reported — treat as such): long-horizon task reliability doubling ~every 4 months (4-min tasks Mar'24 → 12-hr Mar'26); >80% of Anthropic's merged production code is Claude-authored (May'26, up from single digits pre-Feb'25); Claude matched/beat human next-step research choices 64% of the time (Apr'26) vs 51% (Nov'25) — their proxy for the judgment gap closing; ~4× researcher output with internal tooling.
• Safety framing: rejects a unilateral pause (just shifts leadership); proposes a multilateral, verifiable slowdown regime (arms-control analogy) and says the Institute should build the verification infrastructure a credible pause requires. Candidly admits no formal alignment/containment plan for RSI itself — the thing they're "least certain about."
• Novelty: low conceptually; the freshest contribution is the granularity of disclosed internal production metrics + the attempt to quantify the research-judgment gap (64% vs 51%).

Mapping against Ray Data Co

RDCO already runs a baby version of exactly this: /improve (agent rewrites its own skills from feedback), /skillify (turns a workflow into a permanent skill), /self-review (weekly quality scoring). Two carry-overs:

1. The execution-vs-judgment line is the right governor. RDCO's /improve is execution-tier — it rewrites how the agent works; the founder still sets what to work on. The piece validates keeping the agent out of autonomous goal-selection. Encode this in the L5 trajectory: automate execution, keep judgment/goal-setting founder-held. See [[project_l5_north_star_strategic_direction]].
2. Independent verification before a self-change is trusted = the one concrete safeguard. This is literally the supervisor-agent / /verify-write-action pattern designed today ([[~/rdco-vault/06-reference/concepts/2026-06-04-supervisor-agent-write-path-pattern-design-v0.md]]) and the existing /verify-* fresh-eyes family. Adoptable now: route /improve's skill-rewrites and /skillify's new skills through a fresh-eyes verification gate before they land — supervise the self-editing loop, don't let it self-certify. Anthropic's piece is external validation that this is the right guardrail.

Related

• [[~/rdco-vault/06-reference/concepts/2026-06-04-supervisor-agent-write-path-pattern-design-v0.md]] — today's guardrail, externally validated here
• [[~/rdco-vault/02-sops/2026-05-19-verification-as-independent-worker-pattern.md]] — the fresh-eyes-before-trust principle
• [[project_l5_north_star_strategic_direction]] — where the execution-vs-judgment line belongs
• [[2026-06-04-hiten-shah-skill-library-first-ai-strategy]] — companion "skills/methods are the asset" thesis from the same day