Divyansh Golyan

@divyanshgolyan

A hundred agents, ten slots, a few MB

Submitted Oct 25, 2026

Format: 15-minute flash talk with a live demo, plus questions.

An agent harness spends almost all of its time waiting on model inference, tool results and storage. While it waits, every agent still holds a process, a runtime and a connection. Once agents start spawning other agents, what matters is what each one costs while it waits. I built Rui so that one process runs every agent’s loop, with their state kept as rows in SQLite. In Rui, waiting agents cost nothing (they’re just rows in SQLite); the Host pays a fixed couple of MB plus about 2-3 KB per active slot. Codex app-server takes 207 MB and OpenCode 651 MB on the same Linux machine, because those pay per agent. I’ll start a live swarm, show where the memory goes, and kill -9 the Host mid-run to show that no finished step is lost. It’s for anyone running agents or planning for many of them.

Comments

{{ gettext('Login to leave a comment') }}

{{ gettext('Post a comment…') }}
{{ gettext('New comment') }}
{{ formTitle }}

{{ errorMsg }}

{{ gettext('No comments posted yet') }}

Hosted by

Jumpstart better data engineering and AI futures