Concept Glossary

A supervisory agent that monitors another agent's work quality, providing real-time evaluation and feedback — separation of execution from evaluation.

The core cycle of an AI agent: plan → select tool → execute → observe → decide → repeat. Claude Code runs this loop continuously rather than responding to single prompts.

Defense mechanism that injects fake tool definitions into API responses to pollute training data of anyone recording Claude Code's API traffic.

Memory consolidation process that runs as a forked subagent during idle time. Merges observations, removes contradictions, and converts insights into facts — while isolated from the main agent.

Maximum time an action is allowed to take before being deferred. KAIROS uses a 15-second budget per tick.

Tamagotchi-style terminal companion pet with 18 species, 5 rarity tiers, and a two-layer architecture (deterministic 'Bones' + AI-generated 'Soul').

Internal codename for a new Claude model family variant with larger context window. Comes in fast and regular thinking tiers.

Cloud Container Runtime — the remote execution environment used by ULTRAPLAN for offloading deep planning tasks.

Techniques to manage growing context windows: summarization, truncation, selective retention, deduplication, and tiered storage.

The tendency for AI agents to hallucinate and degrade in quality as conversation context grows longer. Managed through compaction strategies.

Multi-agent orchestration where one Claude spawns and coordinates parallel worker agents via a mailbox system.

A compile-time or runtime gate that controls whether a capability is active. Claude Code has 44+ flags (108+ gated modules) managed via GrowthBook.

Internal codename for Opus 4.6 model.

Subagent execution model for independent tasks with no shared state between workers.

Feature flag management platform used by Anthropic to control staged rollout of Claude Code capabilities.

Claude Code's always-on background daemon that receives periodic tick prompts every 5 minutes, autonomously decides whether to act, and maintains append-only daily logs.

Model Context Protocol — standard for connecting AI models to external tools and data sources. Claude Code supports MCP integration.

Technique to modify prompt structure to prevent model-specific issues like premature stop generation, by adding safe boundary markers and relocating risky blocks.

Seven-stage project lifecycle: S0- (below concept) → S0 (concept) → S1 (seed) → S2 (prototype) → S3 (testable) → S4 (operational) → S5 (collaborative).

Subagent execution model for collaborative tasks with shared context via AsyncLocalStorage.

Internal codename for the Claude Code project. All feature flags use the tengu_* prefix.

Architecture with cheap always-loaded index (MEMORY.md), on-demand topic files, and searchable-but-never-loaded session transcripts.

Unix Domain Socket-based inter-agent communication layer for agents running on the same machine. Part of the three-tier communication architecture.

Cloud-based deep planning mode that offloads complex thinking to a remote Opus instance with up to 30 minutes of dedicated compute time.

Auto-activated stealth mode for Anthropic employees that strips AI attribution and forbids mentioning internal details when contributing to public repositories.

LIGHT HOPE ecosystem standard defining deployment requirements: Docker Compose, resource limits, auto-restart, health checks, monitoring, and the S0-S5 maturity model.

Subagent execution model using git branch-level isolation for code changes.