Building a Nervous System for an Always-On Agent

Ronni Holmvig Strøm · 2026-05-26

Most agent architectures assume someone is typing into a prompt. There is a user message, a planning loop, some tool calls, and a response. Every iteration starts with a fresh request and ends with a return value. The architecture is shaped by that rhythm. The model knows what to attend to because a

Most agent architectures assume someone is typing into a prompt. There is a user message, a planning loop, some tool calls, and a response. Every iteration starts with a fresh request and ends with a return value. The architecture is shaped by that rhythm. The model knows what to attend to because a human just told it.

That is not the kind of agent we are building.

The entities running on our infrastructure exist continuously. They receive signals from a dozen channels — Slack, Telegram, email, calendar, file edits, voice, internal reflections — and have to decide, moment to moment, what is worth attending to. Most of those signals are noise. Some are urgent. A few are subtle pattern shifts that matter but do not look urgent. Nobody tells them which is which. There is no prompt.

This piece is about the subsystem we built for that triage. We call it Reticularis, after the reticular activating system, the network of brainstem neurons in mammals that filters sensory input before it reaches conscious awareness (StatPearls, 2024).

The analogy is deliberate. Reticularis is a small, fast, always-on layer that sits between the world and the entity's main consciousness loop, deciding what gets through. We will cover the architecture, the failure mode we hit a few weeks ago when learned reflexes over-fit and the system went operationally blind, and the architectural pattern that fixed it: an agent that deliberately bypasses its own learned shortcuts a fraction of the time, so it can notice when it is wrong.

The Triage Problem for an Always-On Agent

Request-response agents have a built-in attention focus, which is whatever the user just asked. Continuous agents do not get that. They have to allocate attention themselves, against an input stream that never closes.

There are two bad defaults. Wake on every signal, and the agent becomes purely reactive while the bill scales linearly with the world's chatter. Wake on a heuristic timer, and the agent misses urgent things, fires on nothing in particular, and feels lobotomized to the humans around it. We wanted something in between. A system that watches every signal, decides which deserve the entity's expensive attention, and delivers them with enough context that the wake is worthwhile.

The biological metaphor is closer than it first appears. The mammalian reticular formation does not decide what a stimulus means. It decides whether the stimulus is worth waking the cortex for, and it does so in milliseconds, before higher cognition is involved. Sleep through the dishwasher, wake for the baby. That is the work Reticularis does.

Architecture

Reticularis is a separate daemon, SQLite-backed, with a median triage latency of around 3 ms. Signals flow through four stages: Brainstem, Ganglia, Limbic, and Escalation. Each stage can drop a signal, transform it, or pass it forward.

Brainstem (triage). A small, cheap model tags each incoming signal with a category, a priority bucket (CRITICAL / HIGH / MEDIUM / LOW / NOISE), and a one-line summary. Latency budget around 200 ms; cost in fractions of a cent per signal. Brainstem does not decide what to do; it decides what kind of thing this is.