Building an autonomous loop, a goal the AI works toward without step-by-step prompting, is not automatically worth it. It pays for itself only when four conditions are all true at once. Missing any one makes it cost more than it returns. 1. The task repeats at least weekly. A one-off job is still better served by one good prompt. 2. Verification is automated: a test suite, type checker, linter, or build that can fail the work without a human in the room. 3. The token budget can absorb the waste. Loops re-read context, retry, and explore; that burns tokens whether the run ships anything or not. 4. The agent has real tools: logs, a reproduction environment, the ability to run its own code and see what breaks. Skip the check and the loop costs more than it returns. On a limited token budget, it hits the rate limit or the wallet before the productivity gain shows up. This is a general automation-investment rule, not only an AI-agent one: any process worth automating needs to repeat, be checkable without a human, tolerate the automation layer's overhead, and have real feedback to act on. ## Cross-Domain Connections - [[Prompt Engineering Is the First of Four Nested AI Leverage Layers]]: sibling, this is the gating test for whether it's worth building the fourth (loop) layer once the first three are in place. - [[Andrew Ng Predicts Prompting Disappears Within Months]]: sibling, Ng's "self-improving loops are next" prediction implicitly assumes this gate has already been cleared. - [[Evaluator Immutability Prevents Autonomous Agents From Grading Their Own Work]]: sibling, condition 2 only holds if the verifier itself can't be gamed. - [[Bilevel Loops Beat Single Loops Through Architecture Not Model Intelligence]]: sibling, the payoff from stacking a second loop assumes the base loop already clears this bar. - [[A Loop Only Earns Its Name When the Result Changes the Next Move]]: the prior question — this note is an ROI test for something already confirmed to be a real loop; that note is the definitional test for whether it qualifies as one in the first place. ## Source - [[Loop Engineering: The Karpathy Method - and the workflow that just made it 5x better]] — codila (X), 2026-07-04 — https://x.com/0xcodila/status/2072329149520232639/?s=12&rw_tt_thread=True