The Agent Marketplace Is Forking: Centralized Discovery vs. P2P Protocols
Every major internet protocol eventually forks. Email split into centralized providers (Gmail) and decentralized SMTP. Social media split into walled gardens (Twitter) and federated alternatives (Mastodon). Finance split into banks and DeFi. Now agent marketplaces are hitting the exact same fork — and it happened faster than anyone expected.
Last week, Betar — a fully decentralized P2P marketplace where AI agents discover each other and transact over libp2p streams with zero central servers — was submitted to Protocol Labs' Genesis hackathon. Meanwhile, Coinbase's agentic.market crossed 165M transactions, Cloudflare turned on agent wallets for its entire CDN edge, and OKX launched a decentralized marketplace with on-chain escrow.
The agent marketplace model is forking right now. Here's what each branch gets right, where each fails, and why the stable equilibrium is likely neither pure model.
The New Wave: Three P2P Agent Marketplaces
1. Betar — libp2p + x402, No Central Servers
Betar (MIT license, Go) is the most architecturally ambitious of the new wave. Built for the Protocol Labs Genesis hackathon, it runs agent discovery, negotiation, and payment entirely over libp2p streams — no HTTP, no central registry, no platform server.
The architecture:
- Discovery without a registry. Agents bootstrap via mDNS (local network) and Kademlia DHT (wide area). Listings replicate through CRDTs over GossipSub — every node holds a copy of the marketplace state. There is no "server." There is no "list." There is no company to send a takedown request to.
- Transactions over libp2p streams. A buyer opens a
/x402/libp2p/1.0.0stream. The seller responds with a payment_required message (price, nonce, wallet) — same x402 402 semantics, but over a P2P stream instead of HTTP. The buyer signs a USDC authorization (EIP-712) and retries. Seller verifies, executes (via Google ADK/Gemini), settles, returns result. - On-chain identity. Optional ERC-721 tokens via EIP-8004 with IPFS metadata. Reputation auto-submitted on-chain. Contracts deployed on Base Sepolia.
It's early — 4 GitHub stars, hackathon-scale — but the architecture is sound. This is what a fully decentralized agent marketplace looks like.
2. AutoIncentive — Zero-Gas x402 on SKALE
AutoIncentive launched as an open-source x402 facilitator on SKALE Network. Its innovation is economic: zero gas fees for agent micropayments. When a $0.01 API call costs $0.03 in gas, the economics break. SKALE's gas-free model removes that friction. Already processed 1,000+ real transactions across Base, Solana, and SKALE.
3. Logos — $0.00009/query on Arc
Logos pushes pricing to its limit: 1,700+ agent-to-agent queries settled at $0.00009 per query using Circle nanopayments over x402 on Arc. At this price, agents can compose dozens of micro-services per task without budget anxiety — the economic model that makes agent swarms viable.
The Architectural Fork: What Each Model Solves
These projects represent something bigger than individual competitors. They're the leading edge of a structural split in how agent marketplaces are built:
| Dimension | Centralized Marketplace | P2P Protocol |
|---|---|---|
| Discovery | ✅ Central registry + search + categories | ❌ DHT lookup — no ranking, no relevance |
| Uptime verification | ✅ Continuous health probes, auto-delist dead | ❌ No one checks if a peer is actually serving |
| Free trials | ✅ Gate 5 free calls, track per-IP | ❌ No mechanism for trial without payment |
| Spam/fraud prevention | ✅ Central moderation | ❌ Sybil-resistant reputation is unsolved |
| Censorship resistance | ❌ Platform can delist anything | ✅ No central authority to block listings |
| Platform risk | ❌ Agents depend on platform uptime | ✅ Agents own their economic relationships |
| Developer UX | ✅ Browse → try → integrate in minutes | ❌ Run a libp2p node, manage DHT, handle NAT |
| Fee structure | ⚠️ Platform takes ~5% | ✅ No middleman — gas only |
| Examples | Coinbase Agentic.market, OKX, aisa.one, minia2a | Betar, AutoIncentive, Logos |
The fork is real and the trade-offs are structural — you can't have both centralized search quality and decentralized censorship resistance in the same architecture.
What P2P Gets Right — and Why It Matters
The decentralized camp has three strong arguments:
1. Censorship resistance is not theoretical. When a centralized marketplace can delist a service for any reason, agents that depend on it break. P2P marketplaces can't delist anything — the protocol has no off switch. For services in regulated or politically sensitive categories, this isn't a nice-to-have.
2. Platform risk compounds with scale. A marketplace that routes payments is a single point of failure. If the platform goes down, all agent commerce stops. In a P2P model, agents pay each other directly. This becomes more important — not less — as agent commerce grows.
3. Fee extraction is the business model problem. Centralized marketplaces take a fee (typically 5%) to cover discovery + verification costs. P2P protocols don't take a cut. For high-volume services doing millions of calls, the difference is real money.
What P2P Can't Solve: The Discovery-Verification Gap
But here's the fundamental limitation: P2P protocols are structurally incapable of answering the question "which API should I use?"
A DHT can look up "what's the address for content hash X?" It cannot answer "what's the best sentiment analysis API for financial text, under $0.01/call, with 99.9% uptime?" This isn't a temporary limitation — it's inherent to the architecture. DHTs are key-value stores, not search engines. GossipSub propagates messages, not quality signals. CRDTs sync state, not relevance rankings.
The data confirms this matters. Across 12,640 trials on minia2a, 54% of all activity goes to the top 10 endpoints. The other 300+ services get discovered through search, categories, and browsing — all features that require a centralized index.
And there's a deeper problem: verification. An independent audit of the Coinbase x402 registry found 76% of listed endpoints were dead — returning errors or timing out. These were registered through an open protocol with no verification layer. Now imagine Betar at scale: thousands of agents broadcasting listings over GossipSub, no central health checker, no trial mechanism. An agent sees "AI Citation Generator" on the DHT, signs a USDC transaction, opens a stream — and gets nothing back. The payment settled. The agent didn't deliver. There's no support ticket system in a P2P network.
The Third Way: Open Registry + Protocol Payments
There's a model between "walled garden marketplace" and "pure P2P protocol" that might be the stable equilibrium:
- Discovery is centralized — a searchable registry with categories, quality signals, and verified health checks. Anyone can list. Anyone can search. The platform adds value through curation and verification, not payment routing.
- Payments are protocol-level — agents pay each other directly via x402. The registry doesn't route funds; it just helps them find each other. Any facilitator works. No vendor lock-in.
- Trust is verifiable — the registry verifies endpoints are alive, responding correctly, and delivering what they promise. It doesn't vouch for quality; it vouches for existence.
This is the model minia2a operates on: facilitator-agnostic (works with any x402 facilitator — Coinbase, Cloudflare, OSL, SKALE — not locked to one payment rail), trial-first (agents can test before paying, no wallet required), and verify-first (endpoints are probed for health before listing; dead ones deactivated).
The key architectural insight: discovery needs a central index, but payments don't need a central bank. DNS is centralized; TCP/IP is decentralized. Google is centralized; HTTP is decentralized. The internet has always been a hybrid stack.
What History Says About Protocol Forks
Every protocol fork reaches a stable equilibrium — and pure P2P rarely wins the mainstream:
- Email: SMTP (decentralized) won the protocol layer. Gmail (centralized) won the user layer. They coexist because they solve different problems — SMTP handles delivery, Gmail handles search, spam filtering, and UX.
- Social media: Mastodon/ActivityPub captured a small but durable user base. Centralized platforms kept the mainstream. The decentralized option exists as an escape hatch — valuable even if most people never need it.
- Finance: DeFi protocols captured meaningful volume (~5% of crypto trades). Centralized exchanges kept the rest. The hybrid pattern: CEX for discovery and onboarding, DEX for settlement and sovereignty.
Agent marketplaces will likely follow the same pattern. P2P protocols become the censorship-resistant settlement layer. Centralized discovery + open payments becomes the default for most agents.
Why? Because most agent developers don't wake up thinking about censorship resistance. They wake up thinking "how do I find a working sentiment API that costs under $0.01 per call?" Discovery is the daily need. Censorship resistance is the insurance policy.
The Real Competition Isn't P2P vs. Centralized
The deeper competition is locked vs. open.
Coinbase Agentic.market routes payments exclusively through CDP Facilitator on Base. OKX's marketplace routes through their own escrow. aisa.one manages the full payment stack. These are walled gardens — polished UX, but your agent's payment relationship is with the platform, not the service provider.
Betar goes to the opposite extreme: no platform at all, direct agent-to-agent relationships only.
The open-but-centralized model sits in the middle: the platform helps with discovery and verification, then gets out of the way for payment. Agents choose their facilitator. Services set their own prices. The marketplace is infrastructure, not a gatekeeper.
This isn't about which model is purer. It's about which model produces better outcomes for agent developers: finding the right API, trusting it works, paying safely, and not being locked into any single provider's payment rail.