ASM - Agent Service Manifest for Circle x Arc

"Agents shouldn't shop. They should settle." ASM is the first open protocol that gives AI agents structured data to evaluate, compare, and select AI services — then route a sub-cent USDC payment to the winner via Circle Gateway x402 on Arc testnet, in one HTTP call. THE PROBLEM When an autonomous agent needs to call an API — translate text, generate an image, transcribe audio — it picks between providers with zero structured data. Result: blind selection, 3–10× cost overrun, non-reproducible decisions. A data problem, not a model problem. THE SOLUTION (3 LAYERS) 1. DISCOVER — Registry of 70 manifests across 47 taxonomies (LLM, image, video, TTS, embed, GPU, DB). Each declares pricing, quality, SLA, and on-chain payment address. 2. EVALUATE — TOPSIS multi-criteria engine ranks candidates by cost × quality × speed × reliability. 72% top-1 taxonomy accuracy. 3. PAY — Each /api/score call resolves a winner and settles $0.005 USDC directly to that provider's Arc address via Circle Gateway. One endpoint, N recipients, sub-second finality. REQUIREMENTS — ALL MET ✅ Per-action ≤ $0.01: $0.005 avg ✅ 50+ on-chain tx: 50/50 settled, 0 failed, 15 unique recipients ✅ Margin: off-chain authorize + on-chain batch settle. Same 50 on L1 = $25–$250 gas. ~5,000× overhead eliminated. ✅ Stack: Arc Testnet (eip155:5042002) + USDC + Circle Gateway GOOGLE TRACK Gemini 2.5 Flash drives the routing loop via Function Calling. Agent receives a task + 30 taxonomies and emits a structured select_taxonomy_and_score call that triggers Circle x402 settlement. 80% accuracy. PARTNER STACK AI/ML API as Provider 0 (google/gemma-3-27b-it) for reranking and intent parsing; Gemini Native as Provider 1. OpenAI-compatible drop-in. LIVE Demo: asm-arc-circle-2026.vercel.app Code: github.com/calebguo007/asm-arc-circle-2026 On-chain: testnet.arcscan.app/address/0xF5d426D5cdfaeB18Ea2cDec2F7c2CB88eEe6b038 MIT.