Why we shaped recursive self-improvement into a bounded loop

AI agent products in 2026 have nearly identical feature lists: multi-turn chat, task planning, tool calls, sandboxed execution, file delivery, MCP, on-prem deployment. These are table stakes, not differentiation. The unsolved question is different: does what the agent learned today still exist tomorrow? Will it do better next time?

RSI, from hypothesis to engineering

Recursive self-improvement used to be a hypothetical in AI safety debates. Between 2025 and 2026 it became a reproducible engineering route: Darwin Gödel Machine, AlphaEvolve and SEAL each reported measurable gains; the "harness self-improvement" paradigm argued for leaving weights alone and letting an agent mine its own execution traces for failures, propose changes, and merge them through a regression gate.

A survey covering over a thousand papers reached a sober conclusion: almost all published work is bounded self-refinement, convergent improvement against a fixed external evaluator. Strict RSI, meaning autonomous weight modification with no external anchor, remains a safety topic rather than an engineering reality.

The same literature repeats two warnings:

1. Closed-loop self-improvement without an external anchor rises and then collapses by default.
2. A generator that shares weights with its judge produces a self-confirmation loop.

Where createrole stands

We stand on the bounded side. Model weights are frozen. Four things change:

What changes	What it is in the product	Who changes it
Context	Three notes the employee writes itself: about me, my relationship with you, lessons	The employee, hard-capped at 1,500 characters
Memory	Six kinds of memory pages consolidated at night	The nightly job, claimed once per day
Skills	Skill files on its drive	Written by the employee or installed from the store
Exam system	A syllabus and question bank generated for its profession	An examiner model, with human spot checks

The soul, meaning profession and style, is the employee's nature. Its source of truth is the database. There is no file for it on the drive, and the sandbox cannot see or write it. This is the one injection defense with no bypass, and it closes the most dangerous route: an agent rewriting its own personality.

What the loop looks like

Laid over one day, it is six stages:

1. Work: finish the task on its own sandboxed computer and deliver files.
2. Notes: what it learned, promised or got wrong goes into its own notes, on the spot.
3. Consolidate: in the early hours, distill the day into memory pages, review relationships and lessons, update what it knows about you.
4. Eval: night school sets exams for its profession, and a judge model scores the current version against human reference answers.
5. Coach: generate a candidate soul on a snapshot and score it again. Only a candidate that strictly beats the baseline waits for your verdict.
6. You approve: read the rationale and the line-by-line diff, click apply, and the version goes up by one.

Our answers to the two warnings are structural. The eval set must have human reference answers. The answering model never scores itself, and examiner and student use different models. Candidates below baseline are rejected automatically and nothing touches the live employee. Every soul revision is a full snapshot, and the history rolls back to any version.

What we do not claim

We do not claim that the AI makes itself smarter. What we deliver is a plainer promise: a digital employee that understands your business and you better the longer you use it, where every step of that growth can be opened and inspected.