Shopping is moving behind a chat box. Nobody is refereeing the deal.
Buyers are starting to shop through an assistant. Sellers are being asked to expose their catalog, their prices and their payment rails to software they did not write. agentmesh sits in the middle: a marketplace prototype where buyer and seller agents negotiate inside deterministic bounds. The protocol layer includes hash-linked audit primitives; durable per-turn persistence is a production milestone.
Four protocols landed in eighteen months
The pieces an agent needs to buy something arrived in a rush. Tool access, signed authorization, agent-scoped card credentials, checkout inside a chat surface. What did not arrive is the part where two agents with opposing interests settle a price and leave evidence behind.
- Nov 2024MCP opens up tool callingAnthropicAgents get a common way to reach tools and data. This is the plug.9
- Apr 2025Agent PayMastercardAgent-specific tokenized credentials with consumer-set limits.8
- Sep 2025ACP ships in ChatGPTOpenAI + StripeInstant Checkout puts real carts inside a chat window.5
- Sep 2025AP2 mandatesGoogle + 60 partnersCryptographically signed intent and cart mandates. Proof of who authorized what.6
- 2025-2026Card rails adaptVisaIntelligent Commerce and Trusted Agent Protocol extend cards to agent-initiated spend.7
Three circles, drawn honestly
Global ecommerce that an agent could plausibly transact by 2030. Wide because the forecasts are wide: Morgan Stanley models about 10% of US ecommerce in its base case, McKinsey models trillions globally.
- Global agent-mediated GMV, 2030
- $1T-$5T
- US share, base case
- $190B
- US share, bull case
- $385B
Forecasts disagree by an order of magnitude. We plan against the low end.2,3,4,10
Revenue after real settlement is integrated
A future option after live settlement, merchant verification, refunds, and disputes are implemented.
The app records run estimates but the standalone build does not invoice them. Local-model inference stays on the buyer's Mac.
A pricing hypothesis, not an active subscription or promise of production features.
Inference is the main variable cost and it falls every quarter. A scripted negotiation costs fractions of a cent because most turns never reach a model. Buyers who point us at Ollama or LM Studio move that cost onto their own hardware entirely.
The record is the product
Anyone can wire an LLM to a product feed. The hard part is what a merchant asks in the second meeting: what did your agent promise on my behalf, who approved the spend, and can you prove the log was not edited afterwards. This prototype enforces spending caps outside the prompt and includes hash-chain construction and verification code. The interactive demo transcript remains in memory; durable bilateral audit storage is still required before production.
Construction and tamper-verification are tested; durable bilateral storage is not yet connected to every demo turn.
Caps, scopes and walk-away floors are checked in code before a tool runs, not asked for in a system prompt.
Catalog records, bounded HTTPS tools, and read-only MCP calls provide the integration starting point.
AP2-style signed mandates, MCP tools, ACP-style checkout. We speak to whichever wins.
What could sink this
OpenAI and Google both want checkout inside their own surface. Our answer is the long tail and the audit trail, not competing for the top 500 merchants.
Only a small share of shoppers say they would let an agent buy unsupervised. We treat the consent pause as permanent product, not a phase.
Agent-initiated payments will attract rules on disclosure and liability. A signed, replayable record is the cheapest possible compliance posture.
Scripted negotiation paths and local model support both cut our exposure to any one provider.
Where the numbers come from
- 01Adobe Analytics, AI traffic reportApr 2026
- 02Morgan Stanley AlphaWise, agentic commerce outlookNov 2025
- 03McKinsey, agentic commerce2025
- 04EMARKETER, US retail forecastDec 2025
- 05Stripe + OpenAI, Agentic Commerce ProtocolSep 2025
- 06Google, Agent Payments Protocol (AP2)Sep 2025
- 07Visa Intelligent Commerce2025-2026
- 08Mastercard Agent PayApr 2025
- 09Anthropic, Model Context ProtocolNov 2024
- 10Gartner, agentic AI predictions2025
Forecast figures are third-party projections, not our results. agentmesh is a working prototype with test accounts, mock payments and a scripted negotiation engine. Nothing on this page is an offer to sell securities.