Claude Code powers 2-hour autonomous run

This is the kind of demo that matters more than benchmark theater: if an agent can stay useful for two hours, the bottleneck shifts from raw capability to task scoping, trust, and guardrails.

• –Anthropic’s own product page frames Claude Code as an agentic system that reads codebases, edits across files, runs tests, and handles CI failures
• –The big value is persistence across long, messy work, not single-shot code completion
• –Parallel sessions matter because real teams need multiple threads of work moving at once, not one perfect prompt
• –The main risk is over-trust; autonomy is useful only if developers can review diffs, commands, and side effects cleanly
• –For AI developers, this is a signal that terminal-native agents are becoming a serious workflow layer, not a novelty