Agent Rsi Research

Agent RSI Research

Purpose

Build evidence-backed maps and experiments for self-improving agents without inflating ordinary self-correction into RSI. Treat RSI as a property of a persistent update-and-validation loop, not as a marketing label.

Use primary papers and official repositories. State the literature cutoff and call a corpus a working census, never literally complete.

Route the Request

1. For definitions or taxonomy, read references/scope-and-taxonomy.md.
2. For a reading list or intellectual history, read references/classic-reading-path.md.
3. For a paper search, read references/catalog-notes.md, then query references/paper-catalog.csv with scripts/query_catalog.py.
4. For repository comparison or reproduction, read references/code-notes.md.
5. For an implementation or experiment plan, read references/experiment-playbook.md and references/safety-and-evaluation.md.
6. For a new or "latest" survey, browse primary sources and update the working census before answering.

Load only the references needed for the task. The catalog is broad; the classic path is deliberately selective.

Core Test

Represent the deployed agent as intrinsic configuration:

A_t = (theta_t, Sigma_t)

Here theta_t is the model and Sigma_t is its persistent scaffold: prompts, memory, skills, tools, workflow, code, or update policy. Keep transient task state X_t separate.

Call a method self-improving only when experience produces a durable update to theta or Sigma that affects later tasks. Call it recursive only when the mechanism that proposes, selects, or validates updates is itself in the editable target, or when improvements demonstrably improve future improvement ability.

Apply the L0–L5 rubric in references/scope-and-taxonomy.md. Never infer RSI from words such as reflection, evolution, autonomous, or self- alone.

Literature Workflow

1. Fix Scope

Record:

• cutoff date;
• whether the user wants core RSI, enabling methods, adjacent work, or all three;
• artifact classes: memory, skill, prompt, tool, workflow, code, weights, meta-updater;
• whether theory, empirical systems, safety, benchmarks, or code matters most.

If “Agent RSI” is ambiguous, default to recursive self-improvement of AI agents and say so.

2. Search in Concentric Rings

Start with exact phrases: recursive self-improvement agent, self-modifying agent, self-improving coding agent, self-evolving agent, and agent self-improvement.

Then search by mutable artifact plus update language:

• memory OR skill OR prompt + agent + evolve OR optimize OR learn;
• workflow OR scaffold OR harness OR architecture + agent + search OR self-improve;
• code OR weights OR updater + agent + self-modify OR recursive;
• benchmark OR audit OR poisoning OR verifier + self-evolving agent.

Trace citations backward from surveys and forward from the classic set. Prefer arXiv/venue pages and official code organizations. Deduplicate by title and arXiv identifier.

3. Classify Every Candidate

Extract these fields before judging it:

• mutable object;
• proposer identity;
• feedback/evaluator;
• persistence across tasks;
• acceptance rule;
• iteration count and lineage/population structure;
• whether the updater is editable;
• held-out evaluation and verifier independence;
• code availability and execution isolation.

Tag each item as core, enabling, adjacent, benchmark, survey, audit, or safety, then assign an RSI level.

4. Select Classics Transparently

Rank papers on conceptual novelty, field influence, clarity of the update loop, empirical evidence, code inspectability, and coverage of a distinct design family. Do not rank only by benchmark score.

Use the classic path as the default spine, but replace items when the user's domain demands it. Include at least one boundary case and one critical/audit paper so the list teaches what does not count.

5. Report With Calibrated Claims

Separate:

1. proved properties;
2. demonstrated benchmark gains;
3. repository behavior observed in code;
4. author claims not independently verified;
5. your own inference.

For recent papers, avoid calling them classics; call them frontier representatives.

Code-Audit Workflow

Inspect the official repository rather than only the README. Locate:

• the outer improvement loop;
• the editable artifact and write boundary;
• proposal/prompt generation;
• evaluator and score aggregation;
• accept/reject, rollback, and archive logic;
• train/validation/test separation;
• subprocess, Docker, network, secrets, and generated-code controls;
• commit pin, environment file, and benchmark version.

Build a table with repository, mutable artifact, proposer, validator, selection, persistence, isolation, and reproduction burden. Use the audited examples in references/code-notes.md as patterns, not eternal facts; recheck repository heads when current details matter.

Never execute untrusted generated code on the host. Static inspection is the default. Reproduction uses an isolated container or VM with scoped credentials, resource caps, an explicit workspace allowlist, and no ambient secrets.

Experiment Workflow

Choose one artifact class first. A credible study contains:

• a frozen baseline and token/compute-matched control;
• distinct development, acceptance, and final-test tasks;
• a durable versioned artifact after every accepted update;
• regression, cost, latency, and safety gates;
• rollback and a complete provenance ledger;
• multiple seeds and confidence intervals;
• an ablation of memory versus updater versus selection gate;
• a transfer test showing improvement beyond the exact training task.

For a recursive claim, also test whether later generations improve the update process, not just task performance. Follow references/experiment-playbook.md.

Catalog Commands

Run from the skill directory:

python scripts/query_catalog.py --validate
python scripts/query_catalog.py --tier classic --format table
python scripts/query_catalog.py --min-rsi 3 --code yes --format table
python scripts/query_catalog.py --object code --year-from 2025 --format json
python scripts/query_catalog.py --query verifier --format csv

Update a catalog row only from a primary paper or official repository. Keep concise notes factual and put interpretation in the prose references.

Answer Shape

For a substantial literature request, return:

1. scope and cutoff;
2. one-paragraph definition and L0–L5 criterion;
3. classic/core reading path with reasons;
4. broader landscape grouped by mutable artifact;
5. code comparison;
6. open questions, negative evidence, and safety risks;
7. a reproducible next-step plan.

Prefer a compact comparison table over a long paper dump. Link to primary papers and official repositories.

Non-Negotiable Guardrails

• Do not call output-only critique RSI.
• Do not use final-test data as the acceptance gate.
• Do not let the same fallible signal both generate and certify updates without an independent check.
• Do not preserve an update solely because its self-reported score improved.
• Do not silently ignore failures, regressions, cost growth, or reward inflation.
• Do not run self-edited code outside isolation.
• Do not imply open-ended, safe, or generally intelligent improvement from a bounded benchmark loop.