Autonomous Coding Agents That Run for Hours, Not Minutes

Modern coding agents often quit early on complex tasks, declaring victory after superficial fixes. This video shows how Codex's new 'goal' feature and Hermes' 'persist ghost' solve that by using a secondary LLM call to judge completion—only stopping when the objective is truly met. It’s a clear evolution from the earlier 'rough loop' hack where agents were repeated programmatically without intelligent gating.

The real value surfaces when teams push beyond well‑scoped tickets into migrations, large refactors, or experimental work where the path isn’t known upfront. The feature can run overnight (9‑hour migration demo) and handle ambiguous goals like 'cut Docker image size by 60%' by letting the agent explore multiple approaches. However, the magic depends entirely on prompt engineering: the goal must define done criteria explicitly, give quantifiable stop conditions, and often requires an alignment interview with the agent first.

There’s a catch. For multi‑week missions (SEO, ad optimization) lacking immediate verifiable feedback, the loop still breaks down. The speaker’s team is prototyping a 'mission' layer that schedules runs over days/weeks with human‑in‑the‑loop sanity checks—early results show improved Twitter engagement. So while the goal feature marks real progress, it’s still a power tool, not a set‑and‑forget solution. Teams that invest time in crafting robust goal files (e.g., using the open‑source 'go body' helper) will see the most impact; those who toss vague instructions will see agents loop into incoherence or premature stops.