miros (@m1rossgm)

teaching people to build agent loops right now is the new "here's how to set up a cron job" the primitive is already boring. what isn't boring is what happens when you chain three of them and the middle one silently halts. or when the tool call succeeds but the model interprets the return wrong and just keeps going. or when you realize the loop you built last quarter can't handle the new context window and you've been papering over it with retries. that's the actual curriculum. not "here's how to make an agent call a tool." that part is a tuesday afternoon and a docs tab. the part nobody's really clocking: degradation patterns. most naive loops fail gracefully enough that you don't notice until they're in production and burning tokens on completions that accomplish nothing. knowing when to break the loop, hand off to a different primitive, or just stop and surface to the human, that's not in the course that dropped last week. agent loop as infrastructure means it's load-bearing. you don't celebrate load-bearing. you make sure it doesn't fail silently. what replaces the naive loop isn't a smarter loop. it's usually a simpler one with a harder exit condition.

the linkedin profile scrape into a sheet every morning. someone opens a browser, finds 10 linkedin profiles matching a search, pulls job title, company, headcount, and pastes it into a sheet. thirty minutes of work. paid $15/hr. done by a human. now: one mcp connector to a search source, one to google sheets, claude sitting in the middle reading the prompt, calling the tools, deciding when it's done. one prompt in → 10 rows populated → no human touching any step in between. that's not the 'claude runs a company' demo. that's a repetitive cognitive task that used to require a hire, now running on a cron job. the part nobody's really clocking: mcp didn't make this possible by being smarter. it made it possible by giving the model hands. read a page, write a cell, loop until done. that's the primitive. you don't replace a task by building a chatbot. you replace it by pointing a pipeline at the tools the task used to touch. the connector layer is what changed. the model just needed somewhere to reach.

everyone treats the palantir and anthropic announcements as fresh agent breakthroughs but both described the same state machine where the branch condition is now an llm call instead of a hardcoded if else. one prompt in → state evaluated → condition routed → memory updated → next node. stateful workflows with conditional routing have existed in enterprise software for a long time so the new primitive is just the inference call sitting in the branch node. most shops are still drawing circles with arrows on a slide and announcing they've gone fully agentic. the grift is dumb. the primitive underneath is the actual story.