Overview
Anthropic's leaked code reveals Conway, an unannounced always-on AI agent that represents a shift from model competition to platform lock-in through behavioral memory. This analysis exposes how AI companies are moving beyond chat interfaces to capture persistent user behavior, creating switching costs unlike anything we've seen before.
Key Takeaways
- Always-on agents will learn your work patterns - Conway runs continuously, drafting emails and monitoring channels based on learned behavior, not programmed rules
- The real competition is shifting from models to memory persistence - Success will depend on who controls the agent layer that accumulates your behavioral context over time
- Proprietary extensions create platform lock-in despite open standards - While MCP enables portability, Conway's .cnw format creates a walled garden similar to Google Play Services over Android
- Behavioral context becomes the new form of vendor lock-in - Unlike files or data, the accumulated understanding of how you work cannot be easily exported or migrated to competitors
- Choose your persistent agent platform carefully - Your career effectiveness will increasingly depend on which AI ecosystem you commit to, making provider selection a strategic career decision
Topics Covered
- 0:00 - Conway Agent Discovery: Leaked Anthropic code reveals Conway, an unannounced always-on agent separate from Claude chat
- 1:00 - Conway's Architecture: Three core areas (search, chat, system) with extensions, connectors, and automatic triggers
- 2:00 - Real-World Agent Scenario: Conway running overnight, managing emails, Slack, and calendar based on learned patterns
- 4:00 - Anthropic's Platform Strategy: Conway as capstone of 90-day strategy including Claude Code, Co-work, and Marketplace
- 6:00 - Microsoft Comparison: Anthropic speedrunning Microsoft's 15-year platform dominance strategy in 15 months
- 7:00 - MCP vs Conway Extensions: How Conway's proprietary format sits on top of open MCP standard
- 8:30 - Developer Choice Dilemma: Build portable MCP tools or Conway-specific extensions with built-in distribution
- 10:30 - OpenClaw Acquisition Pattern: How Anthropic copies then blocks third-party tools while acquiring key developers
- 12:00 - New Form of Lock-in: Behavioral context lock-in vs traditional file-based switching costs
- 14:00 - Three Eras of AI Competition: From model quality to interface control to persistent memory ownership
- 16:00 - Enterprise Architecture Implications: Choosing between convenient integrated solutions vs owning your agent memory
- 19:00 - Career and Business Impact: How persistent agents affect promotion, team building, and business decisions
- 21:30 - Employee-Employer Dynamics: How behavioral capture shifts power balance and creates new forms of retention