# Prompt Context Harness and Loop Are Separate Agent Control Problems ## Claim An AI agent's behavior is shaped by four distinct control layers — prompt, context, harness, and loop — and each has its own failure mode, so diagnosing agent problems requires identifying which layer is actually broken rather than defaulting to "better prompting." <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds for agentic (multi-step, tool-using) systems specifically; a single-turn chat completion doesn't have a meaningful harness or loop layer to separate out. - The taxonomy is a diagnostic aid, not a claim that the layers are independent in implementation — a harness change can shift what context gets loaded, for instance. - Doesn't establish which layer is most commonly the actual bottleneck in practice; that's an open empirical question this note doesn't resolve. ## Evidence A tutorial on agent engineering separates prompt (instructions), context (relevant information), harness (tools and guardrails), and loop (repeated action until a verifiable stopping condition) as four layers with distinct failure modes — arguing that treating agent problems as a single "prompting" problem misdiagnoses which layer actually needs fixing. ## Links - [[AI & the Future of Knowledge Work MOC]] — the four-layer vocabulary for diagnosing why an agent behaved the way it did - [[3. Resources/Permanent Notes/AI Output Quality Is Bounded by Staged Context Not Prompt Wording]] — because applies-in; this note's "context" layer is that note's entire claim, restated as one piece of a larger taxonomy. - [[3. Resources/Permanent Notes/Orchestration That Terminates by Code Beats Termination by Model Choice]] — because applies-in; this note's "loop" layer is that note's entire claim, restated as one piece of a larger taxonomy. - [[3. Resources/Permanent Notes/Harness Design Can Matter More Than Model Choice for Agent Performance]] — because applies-in; this note's "harness" layer, with that note's evidence. - [[3. Resources/MOCs/AI Agent Safety MOC]] — because applies-in; a general framework for diagnosing agent failures by layer. ## Tests / falsifiers - Take a set of real agent failures and try to sort each one cleanly into exactly one layer — a high rate of failures that genuinely span two layers at once would weaken the taxonomy's diagnostic value. - Check whether practitioners who use this four-layer frame actually fix problems faster than those who default to "rewrite the prompt" — if not, the taxonomy is descriptive but not actionable. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Literature Notes/How AI Agents Work Prompt, Context, Harness and Loop Explained]] — the tutorial this taxonomy is drawn from.