Claude Code
Anthropic‘s terminal-based agentic coding harness, and the most-referenced artifact in this spoke — named across 84 pages before it had one of its own. This page exists to be the thing those mentions point at.
It is written from what the corpus’s own sources say about it, not from product documentation, which
is why it is short on features and long on how it behaves under measurement. Graded stable under the
2026-08-04 convention — not because Claude Code stands still, but because this is a Thing page: its
currency is the currency of the source pages it draws on, and those carry the url: and the volatile
flag. Every specific below is a dated snapshot; the thing to re-read is the source, not this page.
Why it is load-bearing here
The spoke’s central claim is that the frontier model writes the code and the value moved to everything wrapped around it (agentic-coding-harness). Claude Code is the instance most of the corpus’s sources are actually holding when they make that argument — the harness whose conventions other tools adopt, and the one that keeps turning up on both sides of a comparison.
Its reach shows in what other projects do to reach it. superpowers and compound-engineering-plugin ship cross-harness but treat it as a first-class target; agent-scripts keeps one skill library and syncs it into both Codex and Claude Code; agtop parses its JSONL transcripts to report per-session cost. A tool people write adapters for is a platform, whatever it is billed as.
What the corpus documents
Its instruction file became a convention. claude-md — the per-project markdown loaded into every session — is the vendor-tied sibling of the open agents-md convention, doing the same job. That file is also the thing the strongest measurement here actually tested: see below.
It is available as a library. claude-agent-sdk (formerly the Claude Code SDK) exposes “the same tools, agent loop, and context management that power Claude Code, programmable in Python and TypeScript” — the harness unbundled from the terminal, and the clearest artifact of the model-provider-becomes-platform move the synthesis tracks.
Headless mode is what makes fan-out possible. claude -p runs a stateless one-shot session, which
is the primitive under orchestrate-100-agents-claude-code‘s hierarchical orchestrator and much of
what agent-orchestration describes.
It exposes an effort dial. effort-level — how many files, tools and verification steps the agent spends on a request — is a harness control orthogonal to model choice, and the corpus’s cleanest example of the wrapping doing work the model does not.
Its permission model changed hands (2026-08-14). claude-code-auto-mode records auto mode becoming the default for new sessions on Pro, Max and Team plans — the agent approving its own actions instead of stopping for confirmation. The number that justifies it is about the humans it replaces: per-action reviewers caught 13.6% of harmful actions against auto mode’s 89%, across 1,053 paid developers. See agent-permission-model; the residual 11% and the vendor-generated evidence are the open parts.
Its commercial terms are part of the record, and they move: a temporary +50% weekly allowance
(claude-code-weekly-limits-promo) and a paused plan to bill SDK and claude -p usage at API rates
(anthropic-agent-sdk-billing-pause). The second is the more interesting one — the harness-as-product
has a cost-structure problem that the SDK makes visible.
What measurement says about it
This is the part worth having, because Claude Code is one of the few harnesses in this corpus that has been measured by people who do not sell it.
- context-files-ablation-paper ran it as one of two frontier agents across 288 gold-test runs and found context-injection strategy does not measurably move correctness — bounded to ≤10pp on Claude Code, with the no-context arm scoring higher on borderline tasks. So its own signature convention, claude-md, has the best evidence against it of anything in the spoke.
- jetbrains-ponytail-tested used claude-sonnet-5 on it for 80 paired tasks and 251 audited trials, and found an advertised −54%/−20% effect came back at −15.4%/−10.3%.
- cc-thinking-skills ran a length-matched placebo over 39 skills inside it and returned zero robust verdicts.
Read together, these are the corpus’s evidence for the split recorded in the synthesis: the structural layer around this harness changes how it works — what it reads, what it re-runs, what it spends — and leaves whether it succeeds to the model. Nothing here says Claude Code is ineffective; it says the things people bolt onto it are worth less than advertised.
And the practitioner source agrees from the other direction. claude-code-best-practices, written after shipping three projects with it: “the tool does not matter nearly as much as the person holding it.”
Reflexive note
This hub runs on Claude Code, and the superpowers skill system is what these sessions use. Worth stating plainly rather than leaving implicit, since it is a standing reason to read this spoke’s coverage of it with suspicion.
Related
agentic-coding-harness · claude-md · claude-agent-sdk · agents-md · effort-level · agtop · agentskills-spec · superpowers · agent-scripts · orchestrate-100-agents-claude-code · governed-context-claude-code · context-files-ablation-paper · jetbrains-ponytail-tested · cc-thinking-skills · claude-code-best-practices · claude-code-weekly-limits-promo · anthropic-agent-sdk-billing-pause · claude-code-channels-vs-openclaw