Claude Code
Claude Code
Anthropic's official agentic CLI. Not a code completion plugin — it's an autonomous agent that reads, edits, runs, and debugs code in your terminal. Runs claude-sonnet-5 by default; configurable to claude-fable-5 for the highest-complexity tasks. Reached $1B annualized run rate within six months of GA in May 2025.
What It Actually Does
You describe a task. Claude Code reads your codebase, plans edits, executes shell commands, runs tests, and iterates until the task is done or it's stuck. The loop is: understand → plan → edit → run → evaluate → repeat.
Key behaviors:
- Reads files to understand context before touching anything
- Runs tests after edits to verify changes didn't break things
- Asks clarifying questions when ambiguous, doesn't just guess
- Stops and reports when genuinely blocked (rather than hallucinating a fix)
Hook System
Claude Code has a hook system that lets you intercept the agent loop at defined points:
- PreToolUse: Run before any tool call (file read, shell exec, etc.)
- PostToolUse: Run after any tool call completes successfully
- PostToolUseFailure: Run after a tool call fails
- Stop: Run when the agent is about to exit
- SessionStart: Run when a new session begins
- SessionEnd: Run when a session ends
- UserPromptSubmit: Run when the user submits a prompt
- FileChanged: Run when a file is modified
Hooks enable: auto-formatting on file saves, logging all shell commands, blocking dangerous operations, sending notifications when runs complete. This is the mechanism for overnight-runs monitoring.
Context Compaction
Long sessions accumulate context. Claude Code automatically compacts old context when approaching the model's limit — summarizing earlier conversation history to preserve room for current work. You can also trigger compaction manually with /compact [optional focus]. The project-root CLAUDE.md is re-injected after compaction so project instructions are always in context. This enables multi-hour sessions without manual context management.
Parallel Subagents
Claude Code supports spawning parallel subagents for independent sub-tasks. An orchestrator session fans out multiple agent instances to work on different files or components simultaneously, then synthesizes results. Claude Opus 4.8's Dynamic Workflows feature extends this to hundreds of parallel subagents for large-scale operations (codebase migrations, multi-repo changes) from a single session.
Permissionless Execution
By default, Claude Code asks before running shell commands. You can configure it to run specific commands without asking — useful for trusted operations like npm test, git diff, build commands. Configured per-project via .claude/settings.json. On exe-dev VMs, fully permissionless execution is the default use case.
Pricing
No separate subscription — costs are API token costs at standard Anthropic rates. A typical agentic session (30-60 min of active edits) costs $2-15 depending on codebase size and model choice. Prompt caching reduces costs significantly for sessions that repeatedly read the same files.
Strengths
- Best-in-class for autonomous multi-file editing
- Hook system enables robust agentic-workflows customization
- Works well for overnight-runs on exe-dev or similar remote VMs
- Context compaction enables genuinely long sessions
- Integrates with mcp for extending tool access
- Parallel subagent support for large-scale tasks
Weaknesses
- Terminal-only — no GUI (this is a feature for some, limitation for others)
- Cost can add up on very large codebases where every task requires extensive reads
- Less useful for non-code tasks compared to chat interfaces
Related
claude-sonnet-5 · claude-fable-5 · claude-opus-4-8 · mcp · agentic-workflows · overnight-runs · exe-dev · extended-thinking