Community Deep Dive: Architecture, KAIROS, Auto-Dream, Agent Teams

• Source ID: src-20260409-037a8abb6277
• Kind: analysis
• Scope: shared
• Origin: community-analysis
• Raw path: sources/raw/community-deep-dive-architecture-kairos-auto-dream-agent-teams__src-20260409-037a8abb6277.md
• Status: active

Summary

Community analysis document covering Claude Code architecture, internals, and leak analysis. See extracted entities and claims below.

Tags

architecture five-layer tools

Extracted entities

• Auto Mode Classifier (service): Two-layer defense for auto mode: (1) server-side input probe scans tool outputs
• Speculation Engine (service): Pre-runs predicted user commands before Enter is pressed. Forks a background API
• ULTRAPLAN (service): Remote planning system that offloads complex planning to a Cloud Container Runti
• Five-Layer Architecture (concept): Claude Code's clean five-layer separation: (1) Entrypoints (CLI/Desktop/Web/SDK/
• Agent Teams (service): Multi-instance coordination layer with 3 execution models, 13-operation Teammate

Extracted claims

• [fact] Auto mode classifier has 17% false-negative rate on overeager actions
• [lesson] KAIROS and ULTRAPLAN form a two-speed system (fast/slow thinking)
• [fact] Claude Code is a platform runtime, not a thin API wrapper
• [fact] Speculation engine pre-runs commands in filesystem overlay
• [fact] Seven distinct task isolation variants exist in the source

Extracted relations

• Claude Code --[implements]--> Five-Layer Architecture
• Claude Code --[contains]--> Auto Mode Classifier
• Claude Code --[contains]--> ULTRAPLAN
• Claude Code --[contains]--> Speculation Engine
• KAIROS --[uses]--> ULTRAPLAN