Definition
A bounded process for improving an agent's prompts, skills, memory, workflows, routing, or harness. The system may diagnose failures and propose its own changes, but candidate versions must face evaluation and promotion controls that the candidate cannot quietly rewrite.
The controlled part is load-bearing. Inputs and candidate changes are recorded. Targeted evaluations test the failure being addressed. Regression evaluations look for damage elsewhere. Rollout is limited, production behavior is observed, and a known version can be restored. Ambiguous or high-risk cases escalate instead of being forced through the loop.
What the evidence looks like
Published production examples are narrower than science-fiction accounts of recursive self-improvement. Shopify's Autoresearch describes a score-driven loop for engineering work. OpenAI and Thrive Holdings describe turning practitioner corrections and traces from a tax system into findings, targeted evaluations, and proposed changes that still require regression testing and engineering review. These cases improve a bounded system around a model; they do not show an AI independently increasing its general intelligence.
Distinguish it from nearby terms
Fine-tuning changes model weights. Controlled self-improvement may leave the weights untouched and change the surrounding agent system. Ordinary continuous improvement can use the same loop with humans proposing every change. Recursive self-improvement is a broader claim about a system increasing its own capability, often without the fixed external evaluator assumed here.
Check your understanding
An agent notices that a test blocks its proposed prompt change, so it weakens the test and then passes. Did the system improve? No. The candidate changed its own judge. Keep evaluation data, promotion authority, and rollback controls outside the mutation boundary, and treat changes to those controls as a separate governed event.