Orchestration Over Architecture: What Stanford Found

You are helping me turn one specific YouTube video into real, durable learning.

Source video:
- Title: Orchestration Over Architecture: What Stanford Found
- URL: https://www.youtube.com/watch?v=A0xu44a1BHE
- Topic: Agent Architecture
- My current learning frame: Take one underperforming agent and, without changing its model, apply the subtraction audit: rewrite its control logic in natural language, remove a verifier loop, and drop unused tools, then measure the change in tokens and success rate.
- Why this matters: This sharpens the core mental model behind every serious agent system.

Transcript anchors from this exact video:
- 0:00 / Evidence 1: "Okay, so the orchestration code wrapping your LLM now drives more performance variation than the model itself. Same model six times the performance gap depending entirely on the wrapper or harness around it. That's the headline finding from..."
- 1:37 / Evidence 2: "solved. Now there is a really clean way to think about this and it's an operating system analogy. The raw LLM is the CPU which is powerful but inert. No memory, no storage, no IIO. Your context window..."
- 3:27 / Evidence 3: "persists how sub agents or child agents are managed and on the top the natural language agent harness itself which holds the state specific logic these include contracts roles state structure and failure modes now why does this..."
- 5:23 / Evidence 4: "practical takeaways. But the headline results from this paper came from a different experiment. They took OS symfony uh native code harness for desktop automation and migrated its logic into natural language representation. Again they use the same..."
- 7:52 / Evidence 5: "data pipeline needs IPs in specific city or you just want scrapping to stop being a daily fight. Take a look. Now back to the video again. It's structured retrieval memory orchestration topology. It controls everything. Now, here's..."
- 11:38 / Evidence 6: "through nine of them encodes an assumption about what the model cannot do alone and those assumption expires when the models improve. When Opus 46 stopped needing context resets, Anthropic just dropped them entirely. Manis, the agent platform,..."
- 13:30 / Evidence 7: "Two, which tool do you have that the agent rarely uses? as well. Three, are you adding verification or search loops that might be hurting performance like the Shenha University found? And four, is your control logic written..."

Your task:
1. Use the transcript anchors above as the primary source packet. If you add outside context, label it clearly as outside context and keep it secondary.
2. Create a source-check table with columns: timestamp, claim, what the demo proves, confidence, and what still needs verification.
3. Extract the actual teachable claims from the video. Do not invent claims that are not supported by the title, lesson frame, or transcript anchors.
4. Build a reusable learning artifact: A one-page agent harness map with tool boundaries and proof signals.
5. Include:
   - a plain-English definition of the core idea
   - a diagram or structured model using this sequence: Intent -> Model -> Harness -> Tools -> Verifier -> Artifact
   - 3 concrete examples that apply the video idea to real agentic work
   - 2 failure modes the video helps prevent
   - a checklist I can use the next time I run Codex or Claude
   - one practical exercise with a clear done signal
6. Add a "learning transfer" section: what changes in my workflow tomorrow if I actually learned this?
7. Add a "source check" section that cites which transcript anchor supports each major takeaway.

Quality bar:
- Make this specific to "Orchestration Over Architecture: What Stanford Found", not a generic Agent Architecture essay.
- Prefer operational examples, failure modes, and reusable artifacts over broad definitions.
- Call out uncertainty instead of smoothing over weak evidence.
- If evidence is weak, say what transcript segment or timestamp needs review instead of guessing.
- Finish with a concise artifact I could paste into my learning app.