# Prompt Engineering & Loop Engineering, Clearly Explained (Avi Chawla) > [!question] Lingering Questions > - Does automating the outer loop actually preserve accountability once the human's firsthand understanding of the system degrades? > - Reconcile this thread with the Akshay companion thread — same Boris/Peter source, two write-ups worth cross-referencing? > [!tip] Article Ideas > - A short piece contrasting "the inner loop is solved" with "the outer loop is where the real engineering is" as a two-panel explainer for a DevRel audience. > [!link] Cross Reference > - [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]] — H4's "the agent can't check its own work" restates the same rule. > - [[3. Resources/MOCs/AI Agent Safety MOC]] > [!quote] Vocabulary > - *Harness* — everything wrapping the model: the code running tools, managing state, and handling errors. "Agent = Model + Harness." > - *Outer loop* — the loop that decides what the agent works on next, currently usually a human reading turns and writing the next prompt. ## Bib capsule Avi Chawla, "Prompt engineering & loop engineering, clearly explained!" X/Twitter thread, 2026. [[3. Resources/Readwise/Full Document Contents/Tweets/Prompt engineering & loop engineering, clearly explained!]] ## Author's thesis The inner agent loop (model → tool call → context → repeat) is a solved, trivial primitive dating to ReAct; the open engineering problem is the outer loop — currently the human, reading turns and writing the next prompt — and everything the human quietly provides (stop authority, memory, review) has to be rebuilt if that human steps out. ## Key arguments 1. An agent is just a while-loop; ReAct solved this shape in 2022–23, so nobody competes on it anymore. 2. What's unsolved is the loop around that loop — today, that's usually a human. 3. Taking the human out keeps ownership but loses understanding; stop conditions, memory, and review all have to be rebuilt as system components. ## Claim atoms ### The inner agent loop is solved; the outer loop is not > [!quote] ID:rw_h1 > What wasn't solved is the loop around the above loop, and this is what Boris/Peter talked about recently. The pivot of the whole thread — names the actual open problem. ### In the common setup, you are the outer loop > [!quote] ID:rw_h2 > In the most common setup, you are the loop around the loop. Frames the human's current role precisely: write prompt, read turns, write next prompt, repeat. ### Taking yourself out keeps ownership but loses understanding > [!quote] ID:rw_h3 > One big downside of taking yourself out is that you keep the ownership, but would likely lose the understanding. The core tradeoff of automating the outer loop — not covered by the vault's existing maker/checker notes, which are about the check, not about who understands the system. ### The agent can't check its own work > [!quote] ID:rw_h4 > The agent can't be the one to check its own work, since it will pass whatever it wrote. Same rule as [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]], independently stated. ## Permanent candidates - "The agent can't check its own work" duplicates an existing note — connected, not re-promoted. - "Taking yourself out keeps ownership but loses understanding" is a genuinely new, distinct claim — promoted. ## Processing checklist - [ ] Merge/cross-reference with the Akshay companion thread on the same Boris/Peter source once both notes exist - [ ] Recover the missing "Hard part 3: tools" section from the full Akshay breakdown - [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]] — same independent-check rule restated. - [[3. Resources/MOCs/AI Agent Safety MOC]] — outer-loop automation is agent-safety and governance territory.