long-duration use cases

· Zi Wang · 1 min read

Z / Runner 🏃🏻‍♂️

🏃🏻‍♂️ there are legit long-duration cases; web crawling, giant codebases, infra refactors, compliance audits… still think general agent maybe more differentiated when it is directed w/ recurrence as the primary killer use case.

  • 🏃🏻‍♂️ I agree branching is important, but seems like user-prompted DFS + VM snapshots + LLM prefixes is no longer the ceiling anymore. simular agent s3 per your link below shows it is possible to proactively branch out w/o human in the loop (another example, human prompting is the weak link). If you look/trace manus execution path, it does a great job of collapsing / killing bad paths.

  • 🏃🏻‍♂️ simular doens't seem to use vm, looks like they rely mostly on persistent memory.

  • 🏃🏻‍♂️ visual inspection at the UI layer seems to be the superior (more generalizable); but manus continues to fail with long-horizon performance without me having to re-prompt…

Stephen / Basketball 🏀

  • 🏀 yes, scheduled tasks and event responses are perfect as ai subscriptions. but these are web1 features, not ai unlocks. so, while building the complete stack (mobile log scrolling!), new ai models and modalities and reasoning will first unlock new intelligence.

🏀 #general-agent can decision branching be a key feature? manus pauses to prompt users with multiple options. branching can take the best options along the way to the very end (depth-first search), but users can fork the session with alternative options and additional prompts during review. easily and efficiently implemented using vm snapshots and llm prefixes.

  • 🏀 yes, both are possible / good: auto dfs + manual branching; exhaustive search to maximize benchmark. all models already consider and collapse some alternatives, but some options are points of no return: login or anonymous scraping, form data submits, ..

🏀 2026 scope for general ai agents: virtual computers for 1000 subtasks, compacting context for day-long sessions, self-extending uses for 1m tools, formal code for 0 re-prompts, strength ranking for 10 cross-checks, proactive loading for 100 integrations.

  • tidb's 1m database branching, simular's 72.6% agent s (parallel rollouts), anthropic's agent harness, osworld's failure learning, xiaohongshu's 16k tools.
  • 🏀 yes, simular focuses instead on its behavior best-of-n (bbon) with decision branching and model judge.

🏀 _ manus use cases: data scaffolding, highlight reels.

  • 🏀 the question is why + how; send me 10 more sample logs to debug together?