Long-running agents fail at step 95 of 100: system design is back
arpit_bhayani · x · 2026-11-12
Arpit Bhayani pushes back on the claim that system design no longer matters, arguing long-running agents will make it essential again.
He splits agentic apps into two classes: short-running agents can survive with a simple retry loop since each LLM call is isolated, cheap, and stateless. Long-running agents spanning hours are different — retries are costly in API spend, side effects (e.g. duplicate notifications), and wall clock time.
Common production failures he lists: LLM provider rate limits, network timeouts, OOM-killed pod evictions mid-run, pod rotation during deploys, context window overflow, and human-in-the-loop pauses lasting days.
His key point: most of these aren't AI problems — they're classic software engineering problems reappearing in a new system. The solutions are backend fundamentals: checkpointing, idempotency, durable event logs, and resumable workflows.
Takeaway: abstractions change, fundamentals don't. Eventually you ship your "tutorial" to production.
More from coding & agent
- antirez: AI code generation now works when task boundaries are clear — antirez · 2026-11-12
- One-prompt 3D Solar System: Qoder's new Qwen4-Max-Preview builds an explorable browser universe — hey_abusiddik · 2026-11-12
- Giving agents a whiteboard: Claude Code crew plans on a shared Excalidraw canvas for $0.36 — pramodk73 · 2026-11-12
- AI agents automate the grunt work engineers learn from — how to train the next generation? — bendee983 · 2026-11-12
- KEEL, an Open Standard for AI Coding Project Docs, Launches keel.to — DominusFL · 2026-11-12
- PyTorch 2.15 regression slows LU solver up to 100x, warns researcher — ducha_aiki · 2026-11-12