Expert Discusses Running Autonomous Long-Running Coding Agents

@omarsar0 I run these daily, agents coding, compiling, debugging and backtesting futures strategies inside NinjaTrader 8 and MT5 completely on their own. The real unlock was the validation pipeline around them, so I just review finished equity curves instead of babysitting sessions.

@omarsar0 The thing that separates a demo from a long-running agent isn't the model — it's whether state survives a bad step. Durable checkpoints, a verifier between steps, a safe resume path. Treat the loop like flight control: assume faults, design the recovery first.

@omarsar0 There is it ! Thanks for the work @omarsar0 🙏

@omarsar0 smart thing to make it freely available, we need more raw convos out there instead of polished takes

@omarsar0 the hard part isn't the agent - it's the checkpoints. curious what recovery strategy they recommend when the run goes sideways at hour 3

@omarsar0 Watching this now—the key bottleneck for autonomous long-running agents is whether the stack can handle non-deterministic error recovery without manual intervention.

@omarsar0 Long runs need runbooks.

@omarsar0 @dair_ai @sparky_42069

@omarsar0 I really appreciate the video, but do you think you could also make a transcript available as well.

@omarsar0 opus 4.8 and gpt-5.5? sounds like you're just paying for a better way to hallucinate

The real leverage in long-running agents is not the autonomy. It is whether the harness has a cheap, reliable signal for whether a step actually advanced the goal. Most real workflows do not have a compiler-style verifier. The loops that compound are the ones that treat every decision point as something that must produce auditable evidence of its effect before the next iteration runs.

@omarsar0 Recorded and free is the right move for content like this.

Keen to see how far these agents actually run without human hand-holding.

@omarsar0 always the agent topics. is it just coding or does it spill into other agent loops too

@omarsar0 run cron loops every 30 min. worst failure wasn't a crash — a loop ran empty for 3 days. diffs looked fine, just not useful. multi-model routing is right but what i check now: is this loop still earning its tokens.

@omarsar0 tbh "bookmark this" energy usually means ill never watch it

but autonomous coding agents running for hours without dying? thats worth the click

@omarsar0 real talk this recording is gold. autonomous agents are the next frontier but nobody talks about failure recovery loops enough