# Loop Engineering Clearly Explained (Akshay Pachaar) > [!question] Lingering Questions > - Is the "harness matters more than the model" benchmark claim ever going to get a named, checkable source, or does it stay an anecdote? > - Does the maker/checker split still hold when the checker is itself an LLM sharing the maker's blind spots? > [!tip] Article Ideas > - A DevRel talk walking the "four hard parts" (stopping, context hygiene, tool design, a critic) with a live coding-agent demo. > [!link] Cross Reference > - [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]] — H3's "critic that can say no" restates the same rule a third time in this source cluster. > - [[3. Resources/MOCs/AI Agent Safety MOC]] > [!quote] Vocabulary > - *Context rot* — the degradation in model decision quality as a long-running loop's context fills with accumulated junk; feeds a "doom loop" where worse decisions add more noise. > - *Loop engineering* — designing the autonomous cycle that drives an agent toward a goal, as distinct from prompt engineering (the words) or context engineering (everything the model sees). ## Bib capsule Akshay Pachaar, "Loop Engineering Clearly Explained," X/Twitter thread, 2026. [[3. Resources/Readwise/Full Document Contents/Tweets/Loop Engineering Clearly Explained]] ## Author's thesis The agent loop primitive is solved and commoditized; the center of gravity keeps drifting outward through prompt → context → harness → loop engineering, and the real work now is designing the goal as checkable success criteria, the loop's brakes, and a verifier that can say no — then stepping back. ## Key arguments 1. Every serious framework runs roughly the same six-line loop; nobody competes on the while-statement anymore. 2. The harness — not the model — is claimed to now drive most of the performance difference between agent setups (an unsourced but widely repeated claim). 3. Four hard parts recur across implementations: knowing when to stop, keeping context clean, giving the agent usable tools, and putting a critic in the loop that can say no. ## Claim atoms ### Builders write loops instead of prompting agents > [!quote] ID:rw_h1 > "I don't prompt Claude anymore. I have loops that are running. My job is to write loops." Boris Cherny's opening line, anchoring the whole "loop engineering" framing across this source cluster. ### The harness now matters more than the model (contested) > [!quote] ID:rw_h2 > Teams have kept the model fixed, changed only the code around it, and jumped from the middle of a benchmark into the top five. Same brain, different loop. Unsourced anecdote — treat as a claim pending citation, not a verified finding. ### A loop needs a critic that can say no > [!quote] ID:rw_h3 > Designing the loop is half the job, and the other half is putting something in the loop that can say no, like a test, a type check, or a real error. Same rule as [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]], stated a third time independently in this source cluster — convergent evidence for the underlying claim even though no single source is authoritative. ### Define done as success criteria, then step back > [!quote] ID:rw_h4 > Don't tell the model what to do, give it success criteria and watch it go. Karpathy's framing of the verifier-first mindset — "done" is proven, not claimed. ## Permanent candidates - "A loop needs a critic that can say no" duplicates the existing vault claim — connected, not re-promoted (third independent restatement across this source cluster). - "The harness matters more than the model" is contested/unsourced but distinct enough to capture with an explicit scope caveat — promoted with a caution flag. ## Processing checklist - [ ] Find a named, citable benchmark for the "harness > model" claim before treating it as established - [ ] Cross-link this note with the Avi Chawla and h100envy threads on the same underlying Boris/Peter source - [x] Connect into the relevant MOC — [[3. Resources/MOCs/AI Agent Safety MOC]] ## Links - [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]] — third independent restatement of the same maker/checker rule in this source cluster. - [[3. Resources/MOCs/AI Agent Safety MOC]] — verifier design and loop brakes are agent-safety infrastructure.