# Claude Code

#entity

## Summary

Anthropic's agentic coding CLI/harness, cited across most of the corpus as the reference [[harness]] and used by every practitioner in it except [[thorsten-ball]] (who runs [[amp]]). Used interactively and in fire-and-forget / scheduled modes.

## Current Understanding

Claude Code recurs as the concrete example of the "universal agent" pattern: a short system prompt, ~10–40 file/shell/search tools, a ReAct runtime loop, standardized key files (CLAUDE.md, `skills/`), and native scheduling. It supports both [[skills-as-memory|skills]] and MCP — and notably converts MCP servers *into* skills so their tools don't all hit context at once, erasing most MCP downsides ([[evolution-of-agent-tooling]]).

**Practitioner limits** (from the [[2026-07-21-larysa-interview|Larysa interview]]) — the corpus's most critical first-hand read:

- **Built-in memory is judged an anti-feature** by Eugene ("it'd be better if it didn't [exist]"); use skills instead. See [[skills-as-memory]].
- **Connectors overpromise.** The agent begins work before establishing whether an integration is implemented or available for the user's account tier — see [[integration-dead-ends]]. Slack has first-class Anthropic support (paid seat required); Teams is treated as a dead end; Figma editing needs a paid account.
- **Model selection matters within the family.** Eugene runs **4.7** in production and avoids **4.8** as "too proactive." Code output is considered production-safe from ~**4.6** onward, excluding authorization and payments ([[code-as-throwaway]]).
- **Testing surface:** browser/web testing works; a mobile emulator got badly stuck.

## Evidence

- **Allie Miller:** one of "4 Claude surfaces" — the maximum-control one, with native scheduling — [[2026-07-14-gap-between-ai-users-irreversible]].
- **Konstantin:** used (leaked) as a hackathon backend; example harness; turns MCP into skills — [[2026-07-14-skills-based-on-git]].
- **Sebastian/Eugene:** Eugene's custom harness is built on Claude Code; solved the printer protocol in ~30 min on-site — [[2026-07-14-sebastian-eugene-interview]].
- **Theo Browne:** markdown-file service piped to Codex/Claude on a cron — [[2026-07-14-everything-we-knew-about-software-has-changed]].
- **Larysa/Eugene:** memory critique, connector dead-ends, 4.6/4.7/4.8 model selection, emulator failure — [[2026-07-21-larysa-interview]].

## Related Pages

- Concepts: [[harness]], [[skills-as-memory]], [[agentic-loops]], [[context-as-scarce-resource]], [[integration-dead-ends]], [[leave-less-room-for-imagination]]
- Entities: [[eugene]], [[larysa]]
- Related tools: [[hermes]] (skills-first harness), [[amp]] (sandbox-first, vendor-curated — the contrasting harness design: orbs instead of a local working directory, Oracle/Painter/Puck instead of user-authored skills), Codex CLI, OpenClaude (a Claude Code fork), Cursor, Conductor

## Contradictions / Uncertainty

- Sources treat Claude Code as the harness baseline but disagree on governance: personal/custom (Eugene) vs company-managed at scale (Sebastian — see [[enterprise-ai-reality]]).
- Model-version judgements (4.7 good / 4.8 too proactive) are one practitioner's production preference at one point in time, not a benchmark, and will date quickly. Status: tentative.

## Next Questions

- What does a compliant, company-managed Claude Code deployment look like for regulated teams?
- Is there a supported way to make the agent check integration entitlements *before* starting a task?