Agent Harness and Claude Code
The model is the smaller half of an agent. What decides whether it does reliable work is the harness: what it reads before it acts, what checks it after, what it remembers, and what it is allowed to hand to a cheaper model. These pages are the measured part of that, most of them from my own setup. The method behind them is harness-engineering-summary.
The harness
- agent-harness-and-always-on-gnome: how I run my work through Claude Code, the scheduled jobs, and why they need a box that never sleeps
- harness-cost-by-model-tier: a behavioral harness took a small model from 3 to 10 accepted runs and made it cheaper than the big one unharnessed
- tdd-inside-the-agent-loop: agent-run TDD cost 3 to 8 times the tokens with no better tests or design
- claude-code-anatomy: Claude Code reverse-engineered: the seven-part loop, five layers of compaction, deny-first permissions
- commerce-agent-blueprint: one agent with skills beats subagents; guardrails live in the harness, not the prompt
- design-md-spec: design tokens plus their rationale in one file the agent can lint against
- reading-agent-output: the terminal as a reading instrument once the work is reviewing an agent’s stream
Context: what the agent reads
- claude-code-from-zero: one-page start for people who do not use agents yet
- writing-claude-md: how to write instruction files, and what did not work
- skills-standard: the SKILL.md format 20+ agents now share
- agent-patterns-stream2: skills load lazily, MCPs eagerly, and four other patterns
- claude-5-5-effort-and-cost: effort from xhigh to high, always-loaded rules from 37.7 KB to 16 KB
- token-efficient-web-requests: 80% fewer tokens per web page by asking for markdown
- claude-subscription-boundary: where a Max plan stops being usable as infrastructure
Memory
- agent-memory-architecture: 4 memory types, 3 loops, 5 systems compared
- inline-agent-memory: notes inside the code with a prefix grep finds (AI-NOTE, AI-TODO, AI-ASK)
- wiki-feedback-loop: corrections and unanswered questions flowing back into this wiki
- solograph-rebuilt-one-binary: my memory server, from a graph database to one Rust binary over SQLite FTS5
Delegation and cost
- lead-model-delegation-economics: 3,000 sessions where a strong lead with a cheap sidekick cost less and scored higher
- cheap-model-delegation-boundary: what a cheap model can read for you, and the two places it cannot
Evals and structure
- plugin-eval-ablation: only the arm without your plugin measures what the plugin adds
- deepeval-llm-testing: 50+ eval metrics with pytest in CI
- agent-benchmarks: SWE-bench, PAC-1, LongMemEval and what each one scores
- sgr-deep-dive: schema-guided reasoning, schemas as the interface between the LLM and code
- effect-ts-for-agents: when a typed effect system pays for itself in an agent pipeline
- github-top-100-agentic-shift: every repo under 18 months old in GitHub’s top 100 is harness tooling
Tools and frameworks other people built for the same problems: agent-landscape.