Agentic Commerce Pt 2: The Door Is Free, the Spend Isn't

In Part 1 I mapped agentic commerce to understand what the incumbents are building for. The answer: free protocols that route the money back to them. About 94% of the fee pool I modeled for 2030 goes to card networks, issuers and processors.¹ This piece looks at what they've left open, the holes an early-stage startup could fill. That's where I'll be looking for investments.

Part 1 found one hole: evidence for agent errors that aren't fraud. Following what agents spend widens it to agent trust, the risk of what agents do, and adds a second: what agents buy. Incumbents are filling two others for free. The door (charging an agent per request at the edge) is free, and the meter (tracking and billing what agents use) is going free. The spend isn't. At seed or Series A, I'd back what agents buy that a platform can't give away, and agent trust, at prices a realistic exit repays.

Agents mostly look; they rarely pay. In August, HUMAN counted 80% of agentic activity on product and search pages and 2.1% on checkout and payment (company-reported). Per-request money hasn't followed (Figure 1): x402stats counted USD 793,308 of organic x402 payments in the 30 days to September 28th. The money that did follow runs through accounts and monthly bills, as Part 1 found.

My base case through 2027: no edge or bot vendor publishes a take of 5% or more on paid agent requests, and organic x402 stays under USD 10M a month. By 2027-12-31, at least two of the next three priced acquisitions of companies whose main product serves agents, in the two open holes, land at five times capital raised or more. So far it's two of three (Tavily and SGNL yes, Stytch no), or one of two without SGNL, which also governs human access.

The opportunity map

Four holes, ranked: two open, one closing, one already filled. Part 1's pass list (protocol tokens, standalone wallets, API catalogs, new chains) still holds.

Most of the money for what agents buy, USD 350M of 460M, went into Exa at USD 2.2B and Parallel at USD 2B; trust's is still seed and Series A (Figure 2).

What agents buy

Agents spend most of their effort finding things, so I think their money lands here first. Exa says more than 400,000 developers use its search (company-reported). The one priced exit paid (Figure 3): Nebius agreed to buy Tavily, an agent search API, for a reported USD 275M, about 11 times the USD 25M it had raised. One deal is no base rate.

The price of filling this hole has moved. Linkup raised a USD 10M seed led by Gradient in February. Assume a seed investor's stake halves by exit. Then a Tavily-sized sale returns 10x only if the seed post-money was about USD 14M or less: USD 19M at 30% dilution, USD 8M at 70%, before preferences. A USD 10M seed that sells 20% to 25% of a company prices it at USD 40M to 50M, and the same sale returns about 3x.

At that price, a 10x needs a USD 0.8B to 1B exit, not another Tavily. Exa and Parallel are the likeliest buyers of a seed company here, and the likeliest to crush one selling general search.

The risk is the bundle. Perplexity sells its own index to developers. Anthropic bills USD 10 per 1,000 searches inside its API. Neither gives search away; the danger is that search rides inside the model call, where the agent never picks a supplier.

What I'm looking for: specialized data an agent re-buys on every task, priced per call, that no lab or search company gives away.

Agent trust, priced by who buys it

It's the area I'm closest to, so I'd price it hardest. This hole sits in the rules: the IETF's Web Bot Auth charter leaves end-user authentication out of scope, so proving whom an agent acts for is open, and card rules leave non-fraud agent errors unallocated (Part 1's seam).

Enterprise buyers already pay. Cymphony, which controls what enterprise agents can access, says it reached seven figures of annual recurring revenue in its first year of sales (company-reported); Sequoia co-led its USD 25M Series A at more than USD 100M post-money. Acquirers pay too: CrowdStrike agreed to pay about USD 740M for SGNL, which governs access for people and AI agents inside companies. Ribbit led AIUC's USD 40M Series A to certify and insure agents, a business built on the liability Part 1 left open.

Customer-facing trust is thinner. Twilio's 10-K puts Stytch, "an identity platform for AI agents", at USD 104.1M, below the USD 120M it raised in 2021, when it was valued at USD 1B. t54 Labs² pitches itself in Mastercard's Agent Pay for Machines as "a clear evidence layer" for chargebacks and disputes.

On the same half-diluted stake, a 10x needs a seed post-money of about USD 5M against a Stytch-sized exit and USD 37M against an SGNL-sized one. Cymphony's round is already past that: an SGNL-sized sale makes at most about 3.7x on it.

What kills the customer side: free agent verification at the edge (AWS checks signed agents at no extra cost) and bank sales cycles of 12 to 18 months, per Baselayer's CEO. So could the networks' know-your-agent work, which Ant International, Mastercard and Visa began in September.

What I'm looking for: agent controls, insurance and error evidence sold to buyers already paying for security or disputes, early enough that an SGNL-sized sale still returns 10x.

The meter on agent spend

The meter produced the largest exit, and the window is closing. Stripe agreed to buy OpenRouter at a reported USD 7B to 7.5B, more than five times its May valuation (Forbes). Adyen agreed to pay USD 335M for Orb, which bills usage for sellers, 7.6 times what Orb had raised.

Now platforms are giving it away. Ramp's router is free through 2026 and Cloudflare's AI gateway is on all plans. Its Auto Router, like OpenRouter, puts many models behind one API, falls back when a provider fails and weighs cost per request. Unlike OpenRouter, which offers hundreds of models for a 5.5% fee on credits, it grades each request's difficulty and picks from a Cloudflare-managed pool. Vercel's gateway charges no markup and on September 30th started offering Browserbase's search and fetch tools through the same key.

Pantera led Orthogonal's USD 4.3M seed for a single account and invoice across 50+ API vendors; Nevermined takes 1% of settled volume on stablecoin rails and 2% on cards. If their money is a cut of agent spend, they aren't Part 1's API catalogs.

I'd back one only with a processor sale inside two years in view.

Why I'd pass on the door

AWS fills this hole at zero: it lets a site charge an agent as little as USD 0.001 a request and takes no cut. Cloudflare's Monetization Gateway, in closed beta since September 30th, has no price yet. You can't undercut zero. At AWS's floor the rail eats the whole charge: Coinbase's x402 facilitator charges USD 0.001 a settlement after the first 1,000 a month. TollBit, the independent door, counts 7,000+ publishers and discloses no revenue.

The case against

The strongest case against me: the door gets a price. Cloudflare plans for the gateway to be widely available by early 2027, and its Q2 deck pitches "monetizing trusted automated demand". More than 20% of the web sits behind Cloudflare, so a published take there would prove the price for the rest of the web, exactly where an independent door would sell. If it does, I reopen the pass (watch item 1); until then AWS anchors the price at zero. Incumbents also pay real prices for trust: Visa agreed to pay USD 2.4B for BioCatch, a mature, Permira-backed company. If that's the pattern, a seed check in trust is a long hold to a growth-stage exit, and Stytch's 0.87x is the more relevant price. And independents can stand alone: Baselayer books eight figures of revenue without a network buyer (company-reported), though from business verification.

What I'm watching

1. The door's price. A published take of 5% or more on paid agent requests at a major edge or bot vendor by 2027-12-31, and I'm wrong to pass.
2. Per-request payments, Part 1's item 2. Organic x402 above USD 10M a month for two consecutive months before 2027-12-31, with no counterparty above 25% of a month, means agents pay per request at scale.
3. What agents buy. OpenAI or Anthropic selling a standalone search API, or Exa's auto-search list price (USD 0.007 for 10 results) falling below USD 0.0035, by 2027-06-30, and I move off this hole.
4. The meter. Stripe dropping OpenRouter's 5.5% fee, or a second free gateway billing third-party tools, by 2027-12-31, and the meter has gone the way of the door.
5. Trust's price. By 2027-12-31, an independent identity or dispute company discloses USD 10M of annualized agent-payment revenue (Part 1's early signpost), or a customer-facing one sells for three times its funding or more in a deal announced after 2026-10-02. Either, and I'd back customer-facing trust, not only enterprise control. Neither, and Part 1's reading holds: identity is a feature.

A note on the numbers

I calculated the three figures, the 10x arithmetic and the Part 1 corrections; the 20% to 25% seed stake is my assumption. Everything else is someone else's, named and dated. Data freeze: 2026-10-02; Anthropic's price and the router docs were checked on 2026-10-03. Just ask for the companion data and scripts.

References

Primary data

• x402stats (2026-10-02); Coinbase facilitator (accessed 2026-10-02); AWS WAF pricing (2026-09-25); AWS Web Bot Auth (2025-11-21)
• Cloudflare: agent traffic, Monetization Gateway (2026-09-30); Q2 2026 deck (2026-08-06); AI Gateway (2026-04-20), Auto Router (2026-10-02)
• HUMAN Security, April to August 2026
• Twilio 10-K (2026-02-24); Federal Reserve Board (2025-12); IETF charter

Analysis and market

• SiliconANGLE (2026-02-10); Linkup (2026-02-03); Parallel (2026-04-29); Exa (2026-05-20)
• SecurityWeek (2026-01-08); AIUC (2026-09-15)
• TechCrunch (2026-09-09); TechCrunch, 2021 (2021-11-18); Crunchbase News (2026-09-22)
• Mastercard, Agent Pay for Machines (2026-06-10); Ant, Mastercard and Visa (2026-09-09); Visa (2026-08-03)
• Stripe on OpenRouter (2026-08-19); Forbes (2026-09-15); Adyen (2026-06-11); Pantera (2026-06-25)
• Business Standard (2026-03-02); Fortune (2026-09-30); Kansas City Fed (2026-02-25)

Prices and current conditions

• Perplexity Search API (2025-09-25); Anthropic web search (2026-10-03); OpenRouter quickstart (2026-10-03); Ramp Router (2026-08-19); Vercel, AI Gateway and Browserbase tools (2026-09-30); Nevermined (accessed 2026-10-02)

Prior work in this series

• Part 1 (2026-09-27)

Disclosures: I lead capital formation at PL Capital, which has backed t54 Labs, World and Privy. Views are my own and do not represent PL Capital or Protocol Labs. Nothing here is investment advice. Do your own diligence.