Autonomic

Autonomic is the homeostasis controller for the Agent OS. It regulates agent behavior across three pillars -- operational, cognitive, and economic -- ensuring agents remain stable and resource-efficient without constant human oversight.

The biological analog is the autonomic nervous system: always running in the background, maintaining equilibrium, and intervening only when something drifts out of range.

Design principles

Advisory, not authoritarian. Arcan consults Autonomic before taking actions, but failures are non-fatal. If Autonomic is unreachable, the agent continues with an allow-all default. Regulation never blocks the core agent loop.

Pure controller. The controller logic contains no I/O and no side effects. The homeostatic state is a deterministic fold over the event history -- given the same events, you always get the same state projection. This makes the controller fully testable without mocks.

Hysteresis gates. Mode transitions require a metric to exceed a threshold AND remain there for a minimum duration. This prevents rapid oscillation ("flapping") when metrics hover near boundaries.

Architecture

Three pillars of regulation

Operational pillar

Monitors the runtime health of the agent system:

• Error rate -- ratio of failed operations to total operations. Tracked as a rolling window.
• Latency -- P50/P95/P99 response time for LLM provider calls
• Uptime -- continuous running time since last restart
• Resource pressure -- memory and storage consumption

When operational health degrades, Autonomic recommends more conservative behavior: simpler prompts, fewer tool calls, or switching to faster models.

Cognitive pillar

Monitors the quality of the agent's reasoning process:

• Context utilization -- percentage of the context window currently in use (0.0 to 1.0)
• Memory pressure -- rate of context growth relative to available space. Values: normal, elevated, critical
• Decision confidence -- self-assessed confidence scores from recent decisions
• Task completion rate -- ratio of tasks started to tasks completed successfully

High cognitive pressure triggers recommendations to reduce context consumption, summarize earlier conversation segments, or defer to the human for guidance.

Economic pillar

Monitors the agent's financial resources:

• Budget remaining -- credits or funds available for operations
• Burn rate -- current spending rate extrapolated to hourly/daily rates
• Revenue inflow -- income from completed tasks (if Haima is active)
• Runway -- projected time until budget exhaustion at current burn rate

The economic pillar drives the economic mode, which is the most visible output of Autonomic's regulation.

Economic modes

The EconomicMode enum is defined in autonomic-core:

pub enum EconomicMode {
    /// Balance > 2x monthly burn. Full autonomy.
    Sovereign,
    /// 1-2x monthly burn. Prefer cheaper models, limit expensive tools.
    Conserving,
    /// 0-1x monthly burn. Cheapest model only, no expensive tools.
    Hustle,
    /// Balance <= 0. Skip LLM calls, heartbeats only.
    Hibernate,
}

Hysteresis gates

Mode transitions use hysteresis to prevent flapping. Each transition has an entry threshold, an exit threshold, and a minimum hold duration:

                Entry          Exit
Sovereign  ───────── 2.0x ──── (always sovereign above 2x)
    │                             │
    │  Below 2.0x for 5min       │  Above 2.0x for 5min
    ▼                             │
Conserving ───────── 1.0x ──── 2.0x
    │                             │
    │  Below 1.0x for 5min       │  Above 1.0x for 5min
    ▼                             │
Hustle ──────────── 0.0x ──── 1.0x
    │                             │
    │  Below 0.0x for 5min       │  Above 0.0x for 5min
    ▼                             │
Hibernate ────────────────────────

The gap between entry and exit thresholds prevents oscillation when the budget is near a boundary. The 5-minute hold duration ensures that brief spikes or dips do not trigger mode changes.

RuleSet evaluation

The RuleSet is a collection of HomeostaticRule trait objects in autonomic-controller. Each rule is a pure function that evaluates the current homeostatic state and returns a set of recommendations:

pub struct RuleSet {
    rules: Vec<Box<dyn HomeostaticRule>>,
}

Rules are evaluated independently. Their outputs are combined into a GatingProfile that represents the consensus of all rules.

Gating profiles

The primary output of Autonomic is a GatingProfile -- a map of operation types to allow/warn/block decisions:

{
  "profile": "conserving",
  "gates": {
    "expensive_model": "warn",
    "cheap_model": "allow",
    "tool_execution": "allow",
    "web_search": "allow",
    "file_write": "allow",
    "speculative_research": "block",
    "long_context": "warn"
  }
}

Arcan reads this profile during its regulation phase and adjusts its behavior:

The gating profile changes dynamically as the homeostatic state evolves. Each gating decision is recorded in the Lago journal as an autonomic.gate_triggered event.

Running Autonomic

cd autonomic
cargo run -p autonomicd
# Listening on http://localhost:3002

With Lago persistence for durable state:

cargo run -p autonomicd -- --lago-data-dir ./autonomic-data

API endpoints

Integration with Arcan

During the regulation phase of the agent loop (after reconstructing state, before the LLM call), Arcan sends:

GET http://localhost:3002/v1/autonomic/gating

If the request succeeds, the gating profile governs what operations are available for that tick. If it fails, Arcan logs a warning and uses an allow-all default. This ensures the core agent loop is never blocked by Autonomic unavailability.

Event persistence

Autonomic events use EventKind::Custom with an "autonomic." prefix in the Lago journal. This enables forward-compatible persistence -- new event types can be added without schema migrations: