Skip to main content
Settings
Search
Appearance
Theme Mode
About
Theme
Jekyll v3.10.0
Environment Production
Last Build
2026-06-19 10:05 UTC
Repository
year-of-ai/year-of-ai.github.io
Changelog Docs Jekyll
Current Environment Production
Build Time Jun 19, 10:05
Jekyll v3.10.0
Build env (JEKYLL_ENV) production
Quick Links
Page Location
Source Code
View on GitHub Edit in GitHub.dev
Configuration
Site Config _config.yml
Page Info
Layout default
Collection none
Path pages/self-improvement.md
URL /self-improvement/
Theme Skin
SVG Backgrounds
Layer Opacity
0.6
0.04
0.08

Self-Improvement

The routine AI-agent fleet that reviews, updates, fixes, learns, aligns, and evolves the year-of-ai hub and framework — a self-monitoring, continuously improving "model for models."

Estimated reading time: 8 minutes

Edit on GitHub

Self-Improvement — models watching models

The hub grows the knowledge bases (AI Orchestration) and spawns new eras on its own. But growing isn’t enough — a system that writes itself also needs to watch and improve itself. This page identifies the routine agent fleet that does that: a meta-layer whose job is the health and continuous improvement of the growth system itself.

Today the central engine runs only the 3-tier grow tick. The framework’s own self-improvement mechanisms (learn, pollinate, distill, evolve) exist but are dormant — built for the old per-repo model and never wired into the central engine. The fleet below wires those up and fills the gaps, forming a closed loop.

The continuous-improvement loop

OBSERVE ledger- collector DETECT health · pages · secret watch GATE prepublish · fw-review LEARN learn- flywheel ALIGN tick-clock · canon wardens FIX hygiene · rollback EVOLVE tier-ROI · meta-audit GOVERN cost ceiling · kill-switch every cycle feeds the next — learnings make each future tick cheaper & better
Each grow tick feeds telemetry into the loop; what the fleet learns is written back into the canonical framework, so the next tick across all repos improves.

The fleet (25 agents)

Each agent is tagged by origin — net-new (build it), rewire (wire up a dormant framework mechanism), or upgrade (make an existing deterministic tool agentic). Priorities P0 → P2.

Monitor / Observe

Agent Pri Origin What it does
telemetry-ledger-collector P0 upgrade The P0 keystone every learn/cost/monitor agent depends on
fleet-pause-killswitch P0 net-new The single global halt the doctrine is missing
pages-deploy-sentinel P0 net-new Built — hourly read-only check that each member repo’s Pages build succeeded + the live site responds; files one sticky issue. The sole liveness signal a rollback agent keys off
secret-expiry-watch P0 net-new Built — daily single-attempt probe of CLAUDE_CODE_OAUTH_TOKEN + LIFECYCLE_PAT ahead of the growth window; files/auto-closes one issue (catches a dead token before ticks burn)
fleet-health-watch P1 net-new Built — daily read-only digest over the evolution ledger: error rate, cost trend, per-member last-grow, stalls; one auto-closing issue (the DETECT half of the loop)
fleet-cost-governor P2 net-new Cost rail on the orchestrator — pre-flight ceiling + budget writer
lineage-state-report P2 net-new Single human pane of glass; a pure aggregator over the other signals

Review

Agent Pri Origin What it does
prepublish-gate P0 net-new The biggest hole: an inline gate that builds + fact/link/tone-checks a tick before it publishes
framework-pr-reviewer P0 upgrade Safety for the self-modification path; gates every framework auto-merge
injection-surface-auditor P1 net-new Audits the workflow-injection surface (untrusted input into prompts; web-tools-with-write)
published-content-auditor P2 upgrade Audits published content for post-publish drift — link/image rot, stale facts, orphans
content-license-attribution-auditor P2 net-new Checks the copyright/attribution posture of sourced facts across the public sites

Fix

Agent Pri Origin What it does
repo-write-serializer P0 rewire Built (convention) — the repo-write-<repo> concurrency lock grow-lineage holds + the documented framework/policy single-PR groups; future writers join them so no two race a branch
repo-hygiene-warden P1 net-new Enforces the invariant that a year repo holds only content + config + .claude + telemetry
model-id-drift-checker P1 upgrade Keeps the hardcoded model IDs in policy.yml from going stale or deprecated
supply-chain-security-warden P1 net-new Adds the missing Dependabot / CodeQL / action-pinning audit on secret-bearing workflows
tick-rollback-sentinel P2 net-new Bounded, human-initiated withdrawal of a bad publish — the rollback the system lacks

Align

Agent Pri Origin What it does
ledger-tickclock-auditor P1 rewire Validates the tick-clock / Evolution Log against reality via a CI check gate

Update

Agent Pri Origin What it does
docs-warden P1 net-new Built — gates every PR + sweeps main so every change is matched by a doc update on the hub’s own doc surface (ADR-0005)
claude-md-canon-warden P1 rewire Re-syncs each repo’s drifted (stale old-model) CLAUDE.md back to canon
adapter-canon-aligner P2 rewire The backward .claude/-to-canon diff — keeps the framework converging, not fragmenting

Learn

Agent Pri Origin What it does
learn-flywheel P1 rewire Mines a stabilized telemetry window and embeds friction-removing edits into the canonical framework

Evolve

Agent Pri Origin What it does
tier-roi-auditor P2 net-new Makes model-tier selection data-driven instead of fixed-by-intuition
recovery-rehearsal-agent P2 net-new Exercises the re-seed / rebuild path in an isolated scratch org (the untested recovery story)
fleet-meta-auditor P2 net-new Watches the watchers — audits the agent fleet itself, with an external dead-man’s-switch

Defects the design surfaced (worth fixing regardless)

  • No evolution ledger — per-tick telemetry is uploaded then discarded after 14 days; the aggregator (telemetry.yml) is dormant and misconfigured (wrong trigger name + artifact pattern).
  • No supply-chain security — no Dependabot / CodeQL; workflows pin actions to floating tags on secret-bearing jobs.
  • A bypassable review — learn.yml force-merges its own PR.
  • A publish race on each year repo’s main (no shared write lock).
  • A prompt-injection surface — an untrusted dispatch input is interpolated straight into a model prompt.
  • CLAUDE.md drift — never re-synced, so year repos carry a stale old-model copy.

Rollout order

  1. Phase 0 — safety scaffolding (before any mutating agent): the kill-switch, the write serializer, removing the force-merge, and the framework PR reviewer.
  2. Phase 1 — keystone + gates + canaries: the telemetry ledger collector first, then the inline prepublish gate, then the secret / Pages / health watchers.
  3. Phase 2 — the loops: the learn flywheel and the align / fix wardens.
  4. Phase 3 — slow structural, economic & meta: tier-ROI, cost governor, content & license auditors, recovery rehearsal, and the fleet meta-auditor.

Full detail

The complete fleet — every agent’s trigger, cadence, model tier, inputs, outputs, guardrails, and the closed-loop rationale — is recorded in ADR-0003. It is the synthesized output of a 14-agent design workflow (4 architect lenses → synthesis → 3 adversarial critics → finalize).