Laura Shin hosts Noah Levine, partner at A16Z, and Robbie Peterson, junior partner at Dragonfly, to discuss the future of agentic commerce on Unchained.

Both guests recently published pieces on how AI agents will transform commerce, though their visions differ significantly in scope and timeline. The conversation explores technical protocols like MPP and X402, payment infrastructure challenges, and the role of traditional card networks versus blockchain-based solutions.

Key topics include the taxonomy of different agent types, fraud protection mechanisms, and how venture capital perspectives are evolving around value capture in an AI-driven commerce landscape.

Three Categories of AI Agents and Their Commercial Impact

Robbie Peterson categorizes agents into three types: commercial agents deployed within businesses, consumer agents for personal use, and bottom-up autonomous agents that transact independently.

"95% plus of most agentic activity is ultimately going to be these commercial agents" - Robbie, drawing parallels to how 95% of software today is deployed within businesses and governments.

Commercial agents primarily automate tasks without necessarily spending money, like Salesforce agents doing automated sales outreach or finance agents handling routine functions.

Bottom-up agents are truly autonomous, procuring their own resources and potentially interfacing with other agents, representing a qualitatively different category from human-supervised agents.

Consumer Agent Limitations and Preference Discovery

Robbie challenges the common vision of agents autonomously booking complex purchases like Tokyo trips, arguing that "preferences aren't a static thing" and are "revealed in the process of discovery itself."

For consumer purchases like travel, clothing, and groceries, agents lack the full context needed for decision-making, missing qualitative inputs discovered during the shopping process.

The more realistic path involves agents providing options and asking follow-up questions rather than making autonomous purchasing decisions on behalf of consumers.

Payment Rails: Card Networks vs Blockchain Solutions

Card networks excel at risk scoring and authentication but lack instant settlement, while blockchains offer instant settlement but haven't solved fraud protection and risk scoring.

"The same technology that powers Apple Pay with tokens is the exact same technology that the networks like Visa and MasterCard are using to power agentic commerce" - Noah.

Noah predicts card networks will dominate traditional consumer-to-merchant transactions, while stablecoins will serve as "digital cash" for new developer-created merchants that are difficult for traditional processors to underwrite.

Blockchains may win in bottom-up agent scenarios due to being "open and permissionless and unencumbered by regulation," enabling more experimentation than traditional rails constrained by shareholders and institutional counterparties.

Fraud Protection and Micro-Transaction Economics

For micro-transactions of "fractions of a cent or a few cents," fraud becomes less relevant because users can accept wasting small amounts rather than dealing with complex fraud protection.

Noah used Allium's X402 server to spend "30 or 40 cents" for data points to build an entire report, demonstrating how micro-payments enable new business models for developer tools.

Card networks retain advantages for high-value purchases where fraud protection and chargebacks remain critical, but lose relevance for low-value, high-frequency developer transactions.

Protocol Competition: MPP vs X402 and Market Dynamics

MPP offers "sessions" (like opening a tab at a bar with a budget) and integrates with traditional payment rails through Stripe, while X402 provides permissionless ERC-20 transactions without sessions.

Noah sees both protocols as viable: "MPP has the benefit of the Stripe ecosystem behind them" for existing merchants, while X402 appeals to the crypto ecosystem's preference for "open permissionless standards."

Visa CLI represents a different approach as a CLI-based wallet downloadable into OpenClaw or Claude Code, coming "default with card" rather than being a protocol.

Adoption Barriers Beyond Technical Infrastructure

"It is not the Rails that's the bottleneck, and it's actually human sort of social structures" - Robbie, identifying bureaucratic inertia and legal frameworks as bigger obstacles than payment technology.

Government spending represents a massive economic segment unlikely to adopt agents quickly, with "plenty of precedents throughout history" showing slow government technology adoption.

Legal and regulatory systems "weren't engineered for machines" and lack the "protocol upgrade" capability that can fix technical rails, creating persistent adoption friction.

Noah remains more optimistic, comparing current agent skepticism to early internet security concerns, believing people will "optimize for what is the lowest barrier to entry" despite risks.

Value Capture and VC Investment Thesis Evolution

Robbie's The Fat Wallet Thesis argues that "whoever owns the end user ultimately captures the most value," contradicting claims that agents will abstract away front-end interfaces.

"People actually like interfaces" and won't want trading charts abstracted away since "the chart itself governs my decision of whether I want to go long or short" - Robbie.

Value accrues in a "barbell" pattern: front-end user ownership and settlement layer control, where "volume and liquidity begets volume and liquidity."

Noah counters that AI is making front-ends increasingly modular, allowing users to "very easily customize what that front end looks like" while keeping back-end infrastructure static, shifting value to unique back-end capabilities.