Self-improvement

Optional, off by default, and policy-gated. Self-improvement lets a generated agent periodically evaluate and improve itself — without ever bypassing its permission model.

Configure

proagent init --intent "..." --self-improving weekly --improvement-policy supervised
proagent improve schedule monthly
proagent self-improve --schedule weekly   # alias documented in the README
proagent improve status --json

self-improve is an alias for improve (README spelling); both configure an agent-building session — run proagent init first.

Frequencies: daily, weekly, monthly, quarterly, manual.

Policies (safest is the default)

Mode	Behavior
propose	Create proposals; change nothing
supervised	Prepare changes, validate, request human approval
auto	Apply only changes satisfying declared safety constraints

What may improve vs what may not

┌─────────────────────────────────┐
│ permissions · secrets · security│   IMMUTABLE
│ constraints · human approval    │
├─────────────────────────────────┤
│ skills · workflows · context    │   IMPROVABLE
│ tools · agent topology          │
└─────────────────────────────────┘

An improvement proposal that would expand permissions, access new secrets, disable validation or remove human approval is treated as high-risk and requires explicit human approval. The CLI stores the immutable set on every policy.

Lifecycle

schedule/trigger → is improvement needed? ── no ──▶ exit (no model calls)
                        │ yes
                        ▼
        research → evaluate vs baseline → propose
                        │
              [supervised] approval gate
                        ▼
              apply → version bump → record diff

• Evaluation before acceptance: the new version must not regress the baseline suite
• Versioning + rollback: agents are versioned; every applied improvement records an explicit diff
• Improvement memory: .improvements/{proposals,applied,rejected,evaluations}/ — rejected ideas are not re-proposed without new evidence

Triggers beyond schedule

manual (proagent improve now), repeated-failure patterns, dependency changes, tool discovery — triggers combine; scheduling never means always running.