70 agents. That's how many Claudee Code](/tools/claude-code/) deployed on its own after a user asked for a "deep search" in ultracode mode - no additional instructions, no manual orchestration setup.
The user reported that Claude Code organized the work into a 4-phase pipeline, spawning roughly 70 separate AI instances (each handling a slice of the search task in parallel) and sequencing them through distinct stages automatically. Ultracode mode is a Claude Code setting that raises the ceiling on how aggressively the tool can fan out work to sub-agents - separate AI processes running simultaneously so complex tasks finish faster than a single back-and-forth session ever could.
What's notable here isn't just the number. It's that the user didn't write an orchestration script or define the pipeline structure. They made a natural-language request, and Claude Code decided how to break it into phases, how many agents each phase needed, and how to feed results from one stage into the next.
This matches how Anthropic has been building out Claude Code's agentic capabilities - the tool is increasingly designed to treat a high-level goal as a project to manage rather than a prompt to answer. The four-phase structure the model built mirrors what a developer might sketch out manually: discover, gather, analyze, synthesize.
The practical ceiling here is cost. Spinning up 70 agents in a single session consumes a significant number of tokens (the units AI models use to process and generate text - think of it as compute credit). Ultracode mode is opt-in for this reason; it's not something you'd run on a routine task. But for codebase-wide searches, large refactors, or deep research that would otherwise take hours of manual prompting, the tradeoff may be worth it.
Claude Code continues to push further into territory where the user defines the goal and the tool handles the execution architecture. Whether that autonomy is reassuring or unsettling probably depends on how much you trust a model to make good decisions about scope.