Self-improving agents are advancing faster than change control
An eight-month GitHub cohort measures whether agent runtimes can change their own prompts, skills, memory, or harnesses, and whether those changes are evaluated, approved, versioned, and reversible.
Early evidence; the market is not yet confirmed
125 subject-filtered implementations show that agent self-improvement is visible, but control is incomplete. Automatic behavior change appears in 100% of the cohort and behavioral evaluation in 8%, while approval appears in 0.8% and rollback in only 0.8%. The ability to change is forming faster than the ability to govern change.
Market snapshot
Comparable measurements from independent market surfaces.
Implementations and packages
Deduplicated and subject-filtered primary cohort.
Median repository stars
Calculated across repositories with at least one star.
Automatic behavior change
125 items in the subject-filtered cohort.
Rollback and recovery
1 items in the subject-filtered cohort.
Search-match dynamics
Bars show monthly GitHub Search API matches. Detailed feature shares use only repositories whose primary metadata explicitly describes agent self-improvement or self-evolution.
What exists inside the category
One item may contain more than one feature.
Execution evidence and traces
4 · 3.2%
Automatic behavior change
125 · 100%
Behavioral evaluation
10 · 8%
Isolated evaluation
8 · 6.4%
Approval gates
1 · 0.8%
Change versioning and provenance
6 · 4.8%
Rollback and recovery
1 · 0.8%
Representative projects
Cross-source validation
These publications are not part of the primary numeric cohort.