agent scheduler + bitter lessons

· Zi Wang · 1 min read

Z / Runner 🏃🏻‍♂️

🏃🏻‍♂️ langgraph: low-level orchestration framework and runtime for building, managing, and deploying long-running, stateful agents & crewai (role based workflow orchestration & LLM routing (comparison).

🏃🏻‍♂️ n by m (in the context of job scheduling); manus brutal force and treat scheduled tasks as an internal cron jobs. looking at the playback logs, manus doesn't seem to be optimizing the sequnce of jobs to make completion time faster. mcp suppose to solve the ai applications w/ multiple data sources, but given most data plumbing jobs are in the "gray zone" it is unrealistic to assume mcp can solve hostile data extraction tasks.

  • meaning, the end-points won't have a clean & pretty json-rpc. while standard stuff like gdrive, slack, direct access to relationship dbs are non-issue (given most llms have native connector integrations already). where manus does get it right is the bespoke on the fly custom connectors it writes; not sure where you see alterantives / better solutions than status quo?

  • also, if we're serious about yt as the 1st testbed for a general agent, then maybe is useful (such as extracting transcripts, metadata, comments ...). no api key is needed, https://glama.ai/mcp/servers/@trilogy-group/youtube-summarizer-mcp and https://www.pulsemcp.com/servers?q=youtube

  • 🏃🏻‍♂️ the point i was making w/ aaron is yt is tooooooo big of a db corpse for search and watching video is too time consuming.

Stephen / Basketball 🏀

  • 🏀 oh, i can do all / 10x better systems design + tooling. it's that "instructions in english and .md – for plan / tool use / retry.. of agents" leading to only max 90% success rate – that baffles me.

  • 🏀 exactly, no humans will casually prompt and no general agent in 2026 will do these: thread-pool task scheduling (java 26!) and impersonating tls fingerprints (curl-cffi).

  • 🏀 i am already building these (best crawlers; later: schedulers); see "abstractions bottom-up" below. but my doubt is bitter lessons = dumb agents.

  • 🏀 oh, the opposite – for best / max crawling, official or third-party spec / toolkit / connectors will never yield.

🏀 "which are the top 10 youtube channels per usage of manus.im / ai (general agents) prosumers? for b-roll, supercut, short..

  • "a better search engine for all their video transcripts, an image analyzer for all their frame-by-frame scenes, a motion tracker for all their 10-second highlight clips…? that is, given a powerful youtube / video ai agent, which channels would you use it on across 1000 videos?
  • "scans 500 transcripts, identifies every segment discussing fasting + gender, cross-references guests who contradict each other, and generates a citation video that jumps between relevant clips.
  • 🏀 that was obviously true, but this? "which can i code to make 100m first? a perfect youtube ai, or a dumb general agent?"