# I guess we're writing loops now? (Theo) > [!question] Lingering Questions > - Does an orchestrator-spawned reviewer agent satisfy maker/checker independence, or is it the maker grading its own homework? This directly tests [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]] if that note exists in the vault. > - What token/time circuit-breaker design bounds a loop that "goes down the wrong path for longer" instead of failing fast? > [!tip] Article Ideas > - "When not to loop" — a rubric combining production risk and API pricing, for the loop-engineering registry (Theo's explicit position: subscription-plan pricing makes loops survivable; API pricing doesn't). > [!link] Cross Reference > - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — agentic-loop practice as a knowledge-work workflow shift > - Loop Engineering material (The New Stack) if present in the vault — practitioner cross-reference for dynamic vs. static loop shape > [!quote] Vocabulary > - none flagged — the source is a practitioner walkthrough using already-common agentic-coding terms (loop, orchestrator, sub-loop) ## Bibliographic capsule Theo (t3.gg), YouTube video/transcript, no fixed publish date extracted. A practitioner account of running autonomous "agentic loops" for multi-day coding work. ## Author's thesis Loop-based agentic coding works, but only within real bounds: the loop's structure should be generated dynamically from the work rather than forced into a fixed template, review should happen agent-to-agent before a human looks at anything, and none of this is free — token cost, not capability, is what currently limits who can do it. He explicitly rejects the no-oversight-autonomy framing some practitioners use. ## Key arguments (≤3) 1. **Loops should be shaped by the work, not a template** — an overnight run that merged four stacked PRs came from a loop that dynamically spawned sub-loops to match the actual task structure. 2. **Reserve human attention for what agents can't clear** — let an agent review code before a human reads it; humans are wasting their own time otherwise. 3. **Token cost is the binding constraint, not capability** — an 8-hour, 3M-token run to fix three small comments was only survivable on a $200/month subscription plan, not API pricing. ## Claim atoms ### Design loops that build loops, dynamically shaped to the work "I asked the model if I could make this loop and it made a loop that makes sub loops dynamically." The distinctive move versus static loop-engineering: the loop's *shape* is generated to fit the work, not forced into a pre-built sprint structure. ### Let an agent review code before any human reads it "If you are reading the code your agent put out before another agent read it and gave feedback on it, you're wasting your own time." A live instance of maker/checker — with the caveat that the reviewer here is co-spawned by the same orchestrator, which may compromise independence. ### Token cost, not capability, is the binding constraint on loops "If you're paying API prices, you probably shouldn't be doing loops yet." An 8-hour / 3M-token run to fix three small comments — survivable only on flat-rate subscription plans. ### Reject no-oversight autonomy; keep a human gate "I am not claiming the same things people like Boris are claiming where they're writing the loop and now the code is just happening by itself with no oversight. That is stupid." His explicit bound: not for million-user production, at least not yet. ## Permanent candidates - **A loop's structure should be generated from the work's structure** — the strongest, most reusable claim; promoted below. - The token-cost claim and the human-gate claim are important operating constraints but read as scoped practitioner observations (pricing model-dependent, and likely to shift as pricing changes) rather than durable atoms — held as Literature-note texture. ## Processing checklist - [x] Checked existing Permanent notes for overlap — "A Nondeterministic Check Breaks the Loops Stop Condition" is a different claim (stop-condition reliability, not loop shape) - [x] Connect into the relevant MOC — [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] - [ ] Resolve whether an orchestrator-spawned reviewer satisfies genuine maker/checker independence — open question, not yet answered in the vault ## Links - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — agentic-loop workflow practice - [[3. Resources/Readwise/Full Document Contents/Articles/I Guess We're Writing Loops Now]] — source document