Swarm24 — a round-the-clock autonomous engineering swarm
A swarm of Claude Code and Codex CLI agents working on the AgroNutikas repos 24 hours at a time: token telemetry, cost attribution, quota circuit breakers and settings that change without a restart.
Problem
One agent does one thing at a time and needs a human at every decision. If development must run at night too, you need a swarm that decides for itself, measures its own cost and stops itself when the quota runs out.
System
Swarm24 runs 6+6 Claude agents (the ClaudeOnly shape) on the AgroNutikas repos, each in its own worktree with its own PR. The control plane measures tokens (cache read vs cache creation), attributes cost to the task and trips a circuit breaker on work that exhausts the quota. Settings are versioned: every worker loads the new effective configuration at the next cycle boundary, not mid-cycle — a half-landed setting does not fail loudly, it runs the wrong config for a whole day. The dashboard shows what is blocking the swarm right now.
Decisions
Ten small dependent PRs, not one big change — every intermediate state on master must be shippable. Every PR proves one load-bearing assertion with a mutation: the test must fail first. The findings log is an artifact updated at the same URL after every major session.
Metrics
6+6 agents, a 24 h cycle, a 10-PR programme (source: agronutikas-workspace, findings log). The “weeks, not quarters” promised on this site comes from here — from the log, not from a slogan.
Stack
Claude Code + Codex CLI orchestration on Windows, a Node control plane, GitHub PR gates.
Product screens
Screens from the Swarm24 operations panel, captured during the UX audit run on 15 August 2026 — the same run as the screenshot above. The panel is local, so the pipeline does not regenerate these on every release.
