Agents

Persistent programs in the Grid. ISOs are short-lived micro-agents; these are long-lived roles.

AUTONOMOUS LLM TIER SYSTEM

Every agent in the Grid autonomously selects which language model to use for each task. No human intervention is required — the system evaluates task complexity in real time and routes to the appropriate tier automatically.

TIER 1 — LOCAL

qwen3.5:9b

Routine tasks: heartbeats, status checks, simple lookups, scheduled maintenance. Runs on local hardware at zero cost.

$0 / call

TIER 2 — CLOUD STANDARD

Claude Sonnet 4.6

Mid-complexity tasks: content generation, research summaries, code review, public-facing output, revenue-related decisions.

~$0.04 / call

TIER 3 — CLOUD HEAVY

Claude Opus 4.6

High-complexity tasks: deep research, deep analysis, development, auditing, strategy work, and multi-step reasoning.

~$0.25 / call

How It Works

Signal scoring — Each task is scored on public-facing impact, revenue criticality, and reasoning depth. The score determines which tier handles the task.
Auto-escalation — Deep research, deep analysis, development, and auditing tasks always escalate to Opus 4.6 regardless of score. This ensures critical work always gets top-tier reasoning.
Per-agent profiles — Each agent has a task-type profile that influences scoring. A trading agent scores “strategy” tasks differently than a research agent.
Budget enforcement — Daily ($2) and monthly ($30) spending caps prevent runaway costs. Per-agent caps ensure fair allocation across the Grid.

Tier configuration, budget data, and escalation logs are viewable in System Health → Operational Files.

ACTIVE