UltraThink
Gated by feature('ULTRATHINK'). Enables extended thinking — deeper, longer reasoning chains that go beyond the standard thinking budget. When activated, the model spends significantly more compute on complex problems before responding.
Auto Mode, AI Buddy & Proactive Execution
In this article we explore some of Claude Code’s most interesting hidden subsystems: automatic permission handling, an AI coaching companion called “Buddy,” extended thinking modes, a verification agent, and away-mode summaries. All of these are gated behind build-time feature flags and runtime checks, meaning they are not available in the public build.
Auto Permission Mode
Section titled “Auto Permission Mode”Claude Code ships with several permission modes that control how tool-use requests (file edits, shell commands, etc.) are approved:
| Mode | Available | Behavior |
|---|---|---|
default | Public | Prompts the user for every potentially dangerous action |
plan | Public | Only allows read-only tools; no writes or executions |
dontAsk | Public | Allows all edits without asking, but still prompts for shell commands |
acceptEdits | Public | Accepts file edits automatically, prompts for shell |
bypassPermissions | Public | Skips all permission checks (the classic “YOLO mode”) |
auto | Ant-only | Uses a transcript classifier to decide automatically |
bubble | Ant-only | Permissions bubble up to the parent process |
How auto Mode Works
Section titled “How auto Mode Works”The auto permission mode is fundamentally different from bypassPermissions. Instead of blindly accepting everything, it feeds the current conversation transcript into a classifier model that decides whether each tool use is safe.
For complex or ambiguous decisions, the classifier employs a two-stage XML thinking approach: it first reasons about the safety implications in a structured XML format before producing a final allow/deny decision.
Configuration is managed through the tengu_auto_mode_config GrowthBook flag, which accepts three values:
enabled— Auto mode is available and on by defaultopt-in— Auto mode is available but must be explicitly selecteddisabled— Auto mode is hidden entirely
YOLO Classifier
Section titled “YOLO Classifier”Separately, there exists a yoloClassifier.ts module that takes a different approach: instead of classifying the full transcript, it specifically analyzes bash commands to determine whether they are safe to auto-approve. Think of it as a smarter version of bypassPermissions that only applies to shell commands it deems non-destructive.
Bubble Permission Mode
Section titled “Bubble Permission Mode”The bubble permission mode is designed specifically for fork subagents (child agents spawned by the coordinator). When a child agent encounters a permission prompt, instead of handling it locally, the prompt bubbles up to the parent terminal session.
This means the parent agent (or the human operator watching the parent) makes all permission decisions, while child agents focus purely on execution. It is a clean separation of concerns in a multi-agent orchestration setup.
Buddy System
Section titled “Buddy System”One of the more unexpected discoveries in the codebase is a full AI coaching companion system located in src/buddy/.
Key Files
Section titled “Key Files”companion.ts— Core companion logicCompanionSprite.tsx— Visual sprite rendering (yes, it has a visual avatar)sprites.ts— Sprite definitions and animationsuseBuddyNotification.tsx— Notification integration hooks
Observer Modes
Section titled “Observer Modes”The Buddy can operate in two modes:
In backseat mode, the Buddy watches your coding session and provides coaching comments. It observes what you are doing and offers suggestions, warnings, or encouragement without directly intervening. Think of it as a pair programmer who watches over your shoulder.
In skillcoach mode, the Buddy actively teaches you skills. Rather than passively observing, it guides you through techniques and best practices, acting more like a tutor than a spectator.
You can also configure it as both to combine the two modes.
GrowthBook cohort events are tracked under the tengu_backseat_* namespace, suggesting this feature has been through (or is currently in) A/B testing among Anthropic employees.
UltraThink and UltraPlan
Section titled “UltraThink and UltraPlan”UltraPlan
Gated by feature('ULTRAPLAN'). A remote multi-phase detailed planning system. Rather than generating a plan in a single pass, UltraPlan breaks planning into phases with a state machine: needs_input transitions to plan_ready once sufficient context has been gathered.
These features represent Anthropic’s experiments with giving the model more time and structure to think through difficult problems, rather than rushing to an answer.
Verification Agent
Section titled “Verification Agent”The Verification Agent is a built-in agent that automatically verifies code changes after they are made. The key design principle is that it runs with “fresh eyes” — it operates in a context separate from the implementation workers, so it is not biased by the reasoning that led to the changes.
This is particularly powerful in coordinator mode, where implementation is delegated to worker subagents. The verification agent can review the workers’ output independently, catching errors that the implementer might have been blind to.
Away Summary
Section titled “Away Summary”The Away Summary feature detects when the user is away from keyboard and generates summaries of what happened while they were gone. When active, an AFK_MODE_BETA_HEADER is sent to the API to signal that the user is not actively watching.
This is useful in long-running sessions — for example, if you kick off a large refactoring task and come back 30 minutes later, the away summary gives you a concise recap of what was accomplished, what failed, and what decisions were made in your absence.