← Autonomy

Autonomous Levels

The L0–L5 scale inherits the TM Forum AN baseline and extends it with AI-specific dimensions: cost circuit breakers, LLM output validation, model fallback chains, tool-use scoping, eval feedback loop.

Manual

System monitors only — no automated actions.

Human role: All actions performed manually.
Real-world example: A blinking server light. The admin reads logs and reacts.

Inside Arcanada

No active Arcanada components live at L0 — every product passes at least the L1 gate.

Adsessor L0→L4 Legal Arcana L0→L3

Assisted

Executes pre-configured deterministic subtasks (cron, pipeline step). Failures emit raw logs.

Human role: Initiates every significant action; reads logs manually.
Real-world example: A backup cron. Runs on schedule, exits 0 or 1, logs to a file.

Inside Arcanada

Email Agent (cron-driven IMAP triage), Long Term Memory (research phase), ARCA Assistant (planning), Auth Arcana (Phase 0), Datarim CLI (human-driven by design — L1 is target).

Verdicus L1→L4 Datarim L1→L2 Disk Arcana L1→L4 Arcanada Support Center L1→L3 Scrutator L1→L3 Long Term Memory L1→L3 Model Connector L1→L3 Arcanada Status L1→L2

Partial

Closed-loop ops in narrow domain with static rules. Health endpoint, classified errors, post-run status to ops dashboard.

Human role: Supervises and approves; triages alerts.
Real-world example: Kubernetes liveness probe + auto-restart on failure within a fixed retry budget.

Inside Arcanada

Verdicus, Transcribator, Ops Bot, Munera, Support, Scrutator, Model Connector, Agent Dreamer — most of the live ecosystem sits here.

Auth Arcana L2→L4 Transcribator L2→L4 Arcanada Ops Bot L2→L4 Muneral / Munera Arcana L2→L3 Email Agent L2→L3 Arcanada Agent System L2→L3

Conditional

Real-time sensing via structured observability (traces, eval scores). Heartbeat, post-deploy smoke gate, validated credentials, explicit exception hierarchy. Adapts via dynamic policies.

Human role: Validates non-trivial decisions; handles unknown failures.
Real-world example: AWS Auto Scaling driven by CloudWatch — a metric crosses threshold and the system scales out. The decision is policy-driven, the human approves the policy, not each action.

Inside Arcanada

Target level for Munera, Support, Scrutator, Model Connector, Agent Dreamer, Email Agent, ARCA Assistant.

Agent Dreamer L3→L4 ARCA Assistant L3→L3

Highly Autonomous · Self-Healing

Cross-domain self-detection, self-diagnosis, self-recovery. Retry-with-classification, circuit breakers, fallback chains, state checkpointing, recovery audit trail, known-fix mapping, post-recovery verification, hard cost circuit breakers for LLM calls.

Human role: Exception handling only — unknown-unknown failures.
Real-world example: Kubernetes self-healing (pod crash → reschedule), Netflix Chaos Monkey, Spinnaker auto-rollback. Plus, for AI agents: hard cost CB (the $437 OpenAI runaway-loop story is what kills you without one).

Inside Arcanada

Target for Verdicus, Transcribator, Ops Bot (first self-healing pilot), Auth Arcana, Disk Arcana, Agent Dreamer.

Fully Autonomous · Self-Optimizing

Closed-loop across services. Proactive self-improvement: A/B testing prompts and models, dynamic model routing, cost governance with auto-downgrade, eval-driven prompt refinement, knowledge graph evolution.

Human role: Sets intent. No routine interaction.
Real-world example: A self-driving system tuning its own routing policies based on observed traffic patterns and a labeled outcome dataset.

Inside Arcanada

Strategic horizon — not a 2026 deliverable. Candidates: Transcribator (when volume crosses 1000+/day), Model Connector (multi-provider routing). Hard cost CB and labeled eval dataset are non-negotiable prerequisites.

AI-specific dimensions

Five dimensions absent from the original TM Forum baseline. Mandatory from level L2+.

LLM cost governance

Real incident 2026-04-29: $437 OpenAI bill in 8 hours from a runaway loop. Hard cost CB is non-negotiable.

Output schema validation

Prompt injection (OWASP LLM01) and schema drift in LLM output. Zod/JSON schema at L3, corrective retry at L4.

Model fallback chain

Provider outage / deprecation / rate limit demands runtime model swap (Claude → GPT → local). Single-model agents fail the L4 gate.

Tool-use scoping

Agentic file-write and shell-exec are a new attack surface (OWASP LLM06 Excessive Agency). High-risk tools are sandboxed at L4.

Eval / RLHF loop

Self-optimization without labeled eval has no signal. Every decision gets a downstream signal; the meta-loop tunes prompts.