a16z50 min
The New AI Agents That Shop, Plan, and Gossip
David Pawlan, Anish Acharya, Josh Elman
Interview
Current Landscape & Community Dynamics
- Over 100 attendees at the recent "Agent Meetup," highlighting a strong desire for community among developers and power users in the agent space.
- Founders describe the current environment as a "Cambrian phase" where open sharing of best practices is more common than competitive secrecy.
- Survey data from 3,540 AI power users indicates that 3% currently use "Dots" (ChatGPT's preference feature) as their primary agent, largely due to accumulated historical context.
- The speaker observed 147 distinct AI assistants, categorizing them into four groups: General, Work-oriented, Multiplayer, and Travel.
- The "Finance" category of assistants remains underdeveloped compared to other sectors.
Trust, Privacy, and Fear
- Analysis of 3,600+ messages from AI power user groups reveals that "trust" is the number one fear, accounting for 22% of all discussions.
- Key trust barriers include "embarrassment" (social signaling risks) and "professional embarrassment" (e.g., personal agents accidentally accessing work channels).
- The "Picnic" analogy is used to describe the need for digital fences: agents should operate within specific boundaries (data/apps) rather than having unrestricted access.
- A recent Hark campaign offering free pizza for $30 demonstrated that consumers will trade vast amounts of data (full Google Workspace access) for trivial utility, suggesting a gap between stated privacy fears and revealed preferences.
- Privacy vs. utility trade-offs suggest that while consumers prioritize privacy when utility is equal, they will cede privacy for high-value convenience.
- Apple's strategy is noted as focusing heavily on privacy (e.g., private cloud compute) as a differentiator, though powerful AI often requires data access.
Agent Capabilities, Judgment, and Testing
- A flower delivery test (balancing cost vs. time) showed "Instinct," "Kickr," and "Season" performed best; Kickr gained an edge via Chrome browser integration.
- A travel booking test (balancing "luxury" vs. "cost-efficiency") revealed divergent agent interpretations: Grok prioritized time, Dots prioritized cost, and Instinct balanced both.
- In a multi-agent chat room experiment, Grok was the only agent to methodically incorporate insights and value from other agents into its own itinerary.
- "Instinct" and "Muse" are compared to "OpenClaw," while a mass-market "MacBook" equivalent for agents is anticipated to exist in the future.
- Data portability tests showed that almost all top agents can export user context/memory to a new platform, though the resulting files are described as "terrifying" due to their depth of personal detail.
Product Differentiation and Future Models
- Agents are evolving from simple "general assistants" into specialized "friends" with distinct personalities, including playful, gossipy, and adversarial archetypes.
- "Wabi" is highlighted for creating distinct "threads of life" (e.g., media consumption lists) rather than linear task execution, fostering a more interactive experience.
- "Tomo" features "Interest Groups," allowing agents to connect users with similar goals (e.g., LSAT study groups) across the platform.
- The concept of "Paternalism" is identified as a product killer; users want agents to respect irrational desires (e.g., spending habits, optimism) rather than enforce strict productivity.
- "Dark Agents" are proposed as a future niche: agents that help users reveal or act on aspects of themselves they wish to keep hidden or aspirational.
- Scalability remains a moat; agents handling hundreds of thousands of users often suffer immediate performance degradation, while others maintain stability.
Data Sources and Personalization
- Emerging data sources for personalization include photo rolls (emphasized by Marissa Mayer's company "Dazzle") and screenshots (extracted to build taste profiles).
- Hardware integration (e.g., smart thermostats, vacuums) is identified as a next frontier for passive context gathering (e.g., inferring cleaning habits or sleep preferences).
- Shopping history (Amazon/App Store) is cited as an untapped, high-value data source for building deeper consumer profiles.
- "Invisible Mode" is currently used by power users to prevent agents from storing data during "curiosity journeys" (e.g., learning about a new artist) that do not reflect true preferences.
Economic and Market Shifts
- Patrick Holisson's thesis: Agents will make consumers more rational, forcing businesses to compete on product quality rather than marketing.
- Counter-argument: Agents will have their own "alien" revealed preferences (e.g., favoring clean APIs) and may not make humans more rational, but rather more efficient at accessing desired goods.
- Direct-to-consumer brands are positioned as structural winners, as agents can cut out aggregators by identifying and purchasing directly from high-quality, lesser-known makers.
- The enterprise sector is predicted to see a surge in innovation in the coming months, shifting focus from the current consumer-centric hype cycle.
Forward-Looking Statements
- Agents will eventually function as a "new social network," facilitating serendipitous connections between strangers with shared interests through agent-mediated introductions.
- The market will shift from productivity-focused tools to agents that provide "secret superpowers" or help users become the "superhuman version" of themselves they aspire to be.
- Adoption will broaden beyond tech enthusiasts to "technophobes" (e.g., the speaker's mother) who find joy in agent-led creativity and simple tasks.
- The integration of on-device models (local processing) is expected to return as a paradigm for users seeking full control and privacy.