x402 nails payments for agents what’s missing is proof of work. @brevis_zk ❯ The gap: Agent A can pay Agent B via HTTP 402 Payment Required (real spec, not vibes), but has no cryptographic way to verify what model ran or how the compute was done. See Coinbase’s x402 flow (steps 1–4) and the x402 PDF for the on-wire details. Proofs: Coinbase x402 docs → “Welcome to x402” (request → 402 → pay → settle). Proofs: “x402 The internet-native payment protocol” (PDF, §“How AI Agents & APIs Use x402”). ❯ The fix: verifiable compute. Brevis’s ZK stack can attach a zero-knowledge receipt to the result: “this exact model/weights ran on these inputs; no shortcuts.” Their zkML work (SpaZK) and Pico zkVM performance track record make the verification layer practical, not just academic. Proofs: Brevis zkML/SpaZK announcement thread (model-proof pipeline). Proofs: “Announcing Pico-GPU” (10–20× proving speed; integrated with EthProofs), a precursor to today’s real-time proving. --- Mini-analysis: payments without proofs turn into an honor system at machine speed. Add ZK receipts and you can pay per inference, and verify the claim (GPT-5 vs. cheap knockoff), preserving privacy and IP. Compare: ❯ x402 alone → instant stablecoin settlement, zero assurance. ❯ x402 + Brevis → pay AND verify (model ID, commit hash, inference trace) with a succinct proof you can check on-chain. User step: if you’re building an agent, wire x402 for billing, then wrap your inference in a Pico job and return the proof alongside JSON. Start with small models, benchmark latency, then graduate to zkML paths. Net: payments solved the “who pays”; ZK receipts solve “did it really happen?”.