01
The Harness
What are you actually building?
A model is a function from context to tokens. Everything that makes it useful is code somebody else wrote.
7 lessons · 14 checks · 58 min
Why it sits here
First because it reframes what the object of work is. Every later module is a component of a harness: the workflow engine is where it runs, the rule compiler is what it invokes, the audit trail is what it emits. Start anywhere else and you are studying parts without the assembly.
Lessons
- Three meanings of one wordDisambiguate before you argue. The word carries a warning people keep forgetting to hear.6m
- The harness is the productNot a slogan. A 3×3 factorial puts a number on it, and the number is larger than most people expect.10m
- The loop, and who decides it is doneThe stopping condition is the most consequential design choice in the whole construct.8m
- Tools are a prompt surfaceNames, descriptions and schemas enter context verbatim on every request. They are the most attended-to text you write.9m
- Context management, where quality actually livesThe constraint is not that the window is finite. It is that quality degrades continuously well before the limit.9m
- Containment over confirmationA dialog approved 93% of the time is not a control. It is a control-shaped object.9m
- Sub-agents are a context techniqueThe benefit is isolation, not parallelism. The most commonly mis-stated point in the subject.7m
Pre-reading
11 sources · ~4h if you read all of itRanked, not exhaustive. The lessons stand on their own, so treat the first tier as the genuinely load-bearing sources and the rest as depth when a lesson makes you want it.
Read properly
- Building effective agentsAnthropic EngineeringThe taxonomy everyone else is implicitly using. Read the original so you do not inherit a garbled second-hand version.25 min
- Stop Comparing LLM Agents Without Disclosing the HarnessZhang, Wang, Ge, Xu, Hamm, Reddy · arXiv:2605.23950The 3×3 factorial behind module 01. The 7.80× variance ratio and the six-of-nine ranking reversals are the numbers that settle the argument.35 min
- Effective context engineering for AI agentsAnthropic EngineeringContext rot, compaction, just-in-time retrieval, with the mechanism explained rather than asserted.25 min
- How we contain Claude across productsAnthropic EngineeringThe 93% approval figure and the 84% prompt reduction. The empirical foundation for containment over confirmation.30 min
- Inside the Scaffold: A Source-Code Taxonomy of Coding Agent ArchitecturesRombaut · arXiv:2604.03515Thirteen scaffolds read at source level. Corrects the idea that the agent loop is a single shape.40 min
- How the agent loop worksClaude Agent SDK docsThe most implementable prose on loop mechanics. Read it as a specification for your own loop. Pair with the permissions page for the six-step evaluation order.30 min
Skim one section
- Writing effective tools for AI agentsAnthropic EngineeringJump to consolidation, error messages and response format. Those are the parts you apply this week.15 min
- Code execution with MCPAnthropic Engineering150,000 tokens down to 2,000 by writing code against a tool surface instead of calling tools through context. An architectural fork, not an optimisation.15 min
- Don't build multi-agentsWalden Yan, CognitionThe corrective to sub-agent enthusiasm. Read against the multi-agent research post; the disagreement resolves on read-versus-write.12 min
- mini-swe-agentSWE-agent teamA hundred lines, bash only, over 74% on SWE-bench Verified. The fastest cure for over-engineering.10 min