Start

Start With Simple Words

A calm first pass that explains recursive self improvement with short sentences and clear pictures.

6 min readEasyUpdated 2026-09-18

The idea in one paragraph

A system studies its own work and keeps what proves useful. It starts with a small task and a clear goal. It tries one change at a time. It tests the change with care. It keeps the win and removes the loss. The next round begins with a stronger method.

Recursion enters when the method itself improves. The system does not only fix answers. It also fixes prompts, tools, checks, memory, and workflow. Better workflow creates better answers. Better answers suggest better workflow edits. The cycle feeds itself while safety rules stay fixed.

A student story

Think of a student with an answer key. The student writes an answer and checks it. Good habits stay. Weak habits leave. The next test goes a little better.

Now replace the student with a coding agent. Replace homework with small tasks. Replace the answer key with tests and judges. The agent writes code. Tests score the code. A separate judge reviews the change. Good changes stay on the mainline. Weak changes return to the archive with a reason.

What makes it count

One rewrite is only refinement. A full loop needs memory across cycles. A lesson from round one must change round two in a visible way. A score must confirm the gain on fresh checks.

A grounded check matters more than praise. A test that can fail is worth more than a judge that always agrees. A small kept win matters more than a large claimed win. This atlas treats proof as the scarce resource.

Plain words for core terms

TermPlain meaningWhere it lives
TraceA saved record of steps, outputs, and costsRecorder and memory
ScoreA test plus judge result that can be comparedEvaluator
LessonA short note that names pattern, cause, and repairReflector
PatchOne small code change plus one matching testProposer
VerdictAccept or reject with a clear reasonGate and judge
ArchiveA home for rejected tries with reasons keptMemory

Verdict rule in plain code

python
def review_candidate(files, change_count, all_passed, quality_gain):
    if len(files) == 0:
        return ("reject", "empty patch")
    if change_count > 3:
        return ("reject", "change too large")
    if all_passed is False:
        return ("reject", "full suite failed")
    if quality_gain < 0:
        return ("reject", "quality regressed")
    return ("accept", "full suite passed")

Continue reading

Found a gap in this page.Open the GitHub repo to propose a fix