Case Study: Claude Code's Long-Running Agent Design

Claude Code (Feb 2025) is Anthropic’s CLI coding agent. Unlike Cursor’s step-by-step collaboration, the philosophy is: give a goal, run to completion.

Positioning

Cursor Agent:  "fix this bug" → edit → you review → continue
Claude Code:   "refactor auth module" → 50 steps later → "done, tests pass"

Optimized for long autonomous tasks, not pair programming.

Core Architecture

CLAUDE.md: Project Memory

Auto-loaded from repo root:

# Project Instructions

## Stack
- Astro 5 + React + Tailwind 4
- Deploy: Cloudflare Pages
- Package manager: npm

## Conventions
- Trailing slashes on routes
- Components in src/components/
- Tests: vitest — npm test

## Commands
- npm run dev
- npm run build
- npm test

Like Cursor Rules + Skills combined—always in system context. Anthropic recommends documenting stack, commands, and conventions here.

Tool Set

10+ built-in tools (MCP support growing):

Tool	Role
Read	read files (line ranges)
Write	create/overwrite
Edit	exact string replace
Bash	shell
Glob	filename patterns
Grep	content regex
WebFetch	fetch URLs
Task	spawn sub-agents

Long-Task Context Management

1. Auto compression

Steps 1–10:  full detail (explore repo)
Step 11:     compress 1–5 to ~200 token summary
Step 20:     compress 6–15
Step 50:     context = summary chain + last 10 full steps + current task

2. Sub-agent parallelism

Main: "refactor auth module"
  ├── sub 1: analyze existing auth
  ├── sub 2: implement JWT
  └── sub 3: update tests
Main merges results, continues

Avoids one sub-task consuming the entire window.

3. Error recovery

Step N:   npm test → 3 failures
Step N+1: read failures, locate bug
Step N+2: edit fix
Step N+3: npm test → 1 failure
… until green

Harness retries and analyzes—not one-shot give up.

Permission Model

Read        → automatic
Write/Edit  → confirm first time; same session may auto-approve similar
Bash        → confirm each command (unless --dangerously-skip-permissions)

WebFetch    → automatic
npm install → confirm

More autonomous than Cursor on file writes; stricter on shell.

Reported Usage (Anthropic)

• Average task: 15–30 steps
• Top uses: bug fix (40%), features (30%), refactor (20%), tests (10%)
• Efficiency gains: 2–3× simple tasks, 5–10× complex
• Common failures: goal drift on long runs; over-editing unrelated files

Harness Comparison

Dimension	Claude Code	Cursor
Interaction	CLI, long autonomy	IDE, collaborative
Context	compress + sub-agents	index + @ refs
Project memory	CLAUDE.md	Rules + Skills
Tools	built-ins + MCP	MCP-centric
Safety	per-command bash gate	per-operation gate
Best for	well-scoped big tasks	exploratory coding

Takeaways

1. CLAUDE.md pattern—worth it on any agent stack
2. Sub-agents—split work, protect context
3. Compression—mandatory for 30+ steps
4. Bash is the danger zone—gate aggressively

References

• Claude Code docs
• Building Effective Agents
• Claude Code best practices

Summary

Claude Code proves CLI long-horizon agents work. Keys: context compression + sub-agents + CLAUDE.md—so 50-step runs stay on goal.