How to Run Claude Code Agents for 24 Hours

Towards Data Science (Eivind Kjosbakken, 2026-07-20): a practitioner how-to for keeping Claude Code (and Codex) agents working unattended for long stretches. Its premise is the spoke’s verification-is-the-bottleneck thesis stated from the operator’s chair: human code review is now the constraint, so the way to get more out of an agent is to cut the interruptions and let it run. This is the temporal-depth companion to the same author’s orchestrate-100-agents-claude-code (parallel breadth) — where that scales agents across tasks, this scales one agent across time. T3: hands-on and opinionated, no throughput/quality data; the “eliminate all bugs” claim for agentic review is unbacked.

The four moves

• Grant permissions, contained. Enable auto-mode / skip per-step approval, but run inside a sandbox and keep backups (GitHub, Time Machine) so a bad step is recoverable. Autonomy is bought with containment, not trust — the agent-guardrails “safety from task choice + reversibility, not human checkpoints” pattern, and the reason a primitive like cloud-run-sandboxes exists.
• Tell the agent how it knows it’s done. “You have to tell the agent how to know when it’s completed its work” — detailed task descriptions with explicit success criteria, and instructions to self-verify by running tests, taking screenshots, or comparing against a baseline (browser access for visual checks). The feedback signal loop-engineering calls binding, written into the task prompt.
• Agentic code review. Use a second agent (he uses Codex via GitHub or CLI) to review the first agent’s PRs, reserving human review for critical components — the two-agent review move as the throughput unlock, not just a safety net.
• Remote, always-on execution. Run on a secondary/always-on machine so closing the laptop doesn’t interrupt the agent; the trade-off is reduced visibility into what it’s doing.

Where it sits

It’s the practitioner how-to counterpart to cognition-fable5-through-the-night (Cognition/Devin running 8h unattended “through the night” on Fable 5): the vendor case study shows long-horizon autonomy is real; this shows the moves a working developer uses to reach for it. Both land on the same load-bearing point — autonomy is gated by the verification signal and the containment, not the model’s raw ability — which is loop-engineering and agent-guardrails restated as an operating manual. It also extends eivind-kjosbakken‘s two-axis picture of scaling Claude Code: breadth (100+ parallel headless sessions) and duration (one agent for 24h), both bottlenecked on the same thing — review the human can no longer keep up with.

Related

eivind-kjosbakken · orchestrate-100-agents-claude-code · loop-engineering · agent-guardrails · agents-never-do-alone · cognition-fable5-through-the-night · cloud-run-sandboxes · devin · anthropic