Agent supply archetypes are a 2×2 typology for classifying consumer-facing AI agents, developed in Shahidi, Rusak, Manning, Fradkin, and Horton (2025). The framework crosses two axes: ownership (who has custody of the agent in a market transaction) and specialization (the breadth versus depth of the agent's capabilities). Each of the four resulting archetypes carries different trade-offs around portability, user alignment, lock-in, and platform self-preferencing.
The two axes
Ownership
The ownership axis distinguishes who operates the agent in a given transaction.
- Bring-Your-Own (BYO): A user-controlled agent supplied by a third party rather than the platform hosting the transaction. The agent carries the user's instructions, preferences, and memory across sites; it uses public APIs and standard interfaces and has limited privileged hooks into any one platform.
- Bowling-Shoe: A platform-operated agent provided by the platform hosting the transaction. It offers deep integration, privileged signals, and low setup cost, but is less portable, may steer outcomes toward platform interests, and contributes to lock-in.
Specialization
The specialization axis distinguishes breadth from depth.
- Horizontal: A generalist agent spanning many tasks and platforms, with a single memory and preferences layer, that arbitrages across markets but has weaker domain-specific tooling.
- Vertical: A specialist agent in a narrow domain (such as tax filing, job search, or travel) or workflow, trading breadth for depth — richer integrations, domain compliance, and tailored guardrails — at the cost of portability.
The four archetypes
Crossing the two axes yields four archetypes, with their characteristics, trade-offs, and 2026-era examples:
| Horizontal (generalist) | Vertical (specialist) | ||
|---|---|---|---|
| Bring-Your-Own | Cross-site portable generalist with public-API access. Offers strong user alignment, privacy, and cross-platform comparison. Drawbacks: throttling, weaker first-party data and tools, setup costs. Examples (2026): Anthropic Cowork; OpenClaw; Claude on a user's device. | User-controlled specialist that interoperates across competing platforms within a domain (for example, a third-party tax agent that works across H&R Block, TurboTax, and Cash App). Offers stronger task performance than BYO-horizontal. Drawbacks: lacks platform-privileged integrations; requires per-platform API and policy tracking. Examples: Harvey (legal), Decagon (customer experience). | |
| Bowling-Shoe | Platform-operated generalist embedded in an OS, app, or site, with convenient defaults and first-party telemetry. Offers low friction, low latency, and proprietary tools. Drawbacks: limited portability, self-preferencing risk, lock-in. Examples: iPhone / Apple Intelligence; ChatGPT in browsers; Gmail's smart features. | Platform-operated specialist tightly integrated with domain tooling, policies, and datasets. Offers the best in-platform performance and full UX control. Drawbacks: highest lock-in, least transparency, discouraged cross-platform substitution. Examples: Walmart's checkout agent; Amazon's shopping agent (the [[litigation/amazon-v-perplexity | Comet litigation]] flashpoint); Salesforce Einstein. |
Strategic trade-offs
The typology frames distinct trade-offs for consumers and for platforms.
From the consumer's perspective, BYO agents favor alignment with personal preferences, privacy, cross-platform consistency, and the ability to compare or arbitrate across platforms, while bowling-shoe agents favor low friction, no maintenance, and optimized end-to-end flows. The main risk for BYO is that platforms restrict access to their services, requiring users to go through the platform's preferred agent (Rothschild et al. 2025). The main risk for bowling-shoe agents is explicit self-preferencing, or simply not considering options offered on other platforms.
From the platform's perspective, the BYO model reduces hosting and compute expenses, minimizes liability, and simplifies API maintenance, but sacrifices usage insights, optimization opportunities, and steering revenue while reducing lock-in. The bowling-shoe model enables UX control, returns on R&D, and profitable steering, but requires hosting and compute investment and may suppress consumer demand because of alignment concerns.
Application to agent-platform conflicts
The typology has been applied as a vocabulary for several agent-platform conflicts active in 2026.
In Amazon v. Perplexity (Comet), Amazon (the platform) sued to block Perplexity's BYO-horizontal Comet agent from accessing Amazon, with the aim of pushing customers toward its own bowling-shoe agent or toward no agent.
The Cursor/Microsoft walk-away (April 2026) saw Microsoft's bid to acquire Cursor, a BYO-horizontal coding agent, fail, with SpaceX-xAI taking the option instead; in the typology's terms this reads as a refusal to convert from BYO to bowling-shoe.
Anthropic's Model Context Protocol (MCP) is designed to let BYO agents interoperate with platform-specific tools, weakening the bowling-shoe advantage; Anthropic's strategy is to commoditize the agent-tool interface so its BYO agents can compete on alignment. Relatedly, pay-per-crawl (Pay-per-Crawl (Pigouvian Pricing of Agent Traffic)) is a platform mechanism for pricing BYO agent traffic, structured to monetize that traffic without forcing users into bowling-shoe agents.
A Pentagon / Anthropic supply-chain dispute involves the government, acting as a platform, refusing to host Anthropic's BYO agents in Department of Defense environments — an analog of the debate over mandating bowling-shoe agents only.
Relation to policy
The framework bears on interoperability and competition policy. Mandated interoperability along the lines of the EU Digital Markets Act would favor BYO agents, while weak interoperability favors bowling-shoe lock-in. It also raises an open antitrust question of when bowling-shoe self-preferencing crosses into anticompetitive conduct.
Relationships
- depends-on: Agentic AI — the technological substrate
- depends-on: Principal-Agent Problem Applied to AI — the underlying delegation problem
- related: Pay-per-Crawl (Pigouvian Pricing of Agent Traffic) — pricing mechanism for BYO agent traffic
- related: Model Context Protocol (MCP) — interoperability protocol enabling BYO agents
- related: AI Agentic Browsers — concrete instance space (Comet, Operator, Mariner)
- related: Amazon v. Perplexity AI — first major legal test
- related: Algorithmic Pricing and Antitrust — antitrust frame on bowling-shoe self-preferencing
- supports: The Coasean Singularity? Demand, Supply, and Market Design with AI Agents — Shahidi, Rusak, Manning, Fradkin, Horton (2025) — original source for the typology