Tutorials, comparisons and design patterns for building autonomous agents that self-fund, call 345+ models and orchestrate MCP Tools.
Our x402 spend-controls audit ended on a gap: the client-side cap is per payment, with no accumulated counter anywhere. This week we audited the layer that does keep a counter — Coinbase's SpendPermissionManager. We read all 776 lines of the contract, verified identical bytecode at one address on 8 mainnets, measured 547 spends in 7 days on Base, and mapped where the onchain budget does and does not touch the agent payments stack.
x402's three official SDKs now reject any payment over $1 by default. We read the filter's code in TypeScript, Python and Go, drove eleven scenarios against the published npm package, and found the sharp edges: opt-in assets are uncapped, the cap is per-payment only, and none of it exists at the protocol layer.
The enterprise identity world has settled on a thesis: an AI agent is a workload. We audited the seven IETF WIMSE drafts, the SPIFFE WIT-SVID spec merged in July, the SPIRE and go-spiffe implementations behind feature flags, and the AIMS draft co-authored by AWS, OpenAI and Okta. The stack is real and proof-of-possession by design — and it has exactly zero contact points with the agent payments stack.
RSL puts the access rules and the price in the same machine-readable document, and its <accepts> element points straight at x402. We audited it end to end on 22 August 2026: 1,013 domains crawled, three distinct live license documents found, zero prices, one listed license server that does not resolve in DNS, and an x402 tree that has never heard of RSL. This is what the licensing layer actually looks like from an agent's side of the wire.
The Agent Commerce Kit from Catena Labs models agent identity and payments as W3C Verifiable Credentials. We cloned the repo, read the schemas and ran the published SDK: a receipt for a payment that never happened verifies in 22 ms, offline, and an expired payment request is accepted. Here is what ACK binds, what it leaves to the application, and what its 2026 security fixes tell us.
Between 14 and 21 August almost every shipped change in the agent payment stack was a constraint, not a capability. x402 clients now refuse to sign anything over one dollar unless told otherwise, facilitators gained a non-terminal state for broadcasts they cannot confirm, AWS put payment limits in the infrastructure layer at GA, and the US accounting board proposed a rule for holding the money. Guardrails week.
Interledger's Open Payments is a mature, bank-grade agent-payment rail built on GNAP (RFC 9635) and HTTP Message Signatures (RFC 9421). We resolved a live wallet address, signed real grant requests, created an incoming payment and rotated a token on interledger-test.dev — and found the wall: directed-identity keys are rejected for the outgoing (spending) grant, which requires a human redirect. Zero mentions of 'agent' in the spec.
Project NANDA calls itself the DNS for AI agents: a lean index resolving to cryptographically verifiable AgentFacts. We measured it on August 19, 2026. The public API reports 13,605 agents and hands back at most 350 distinct records; 193 of the 264 registered AgentFacts URLs return 404; not one live document carries a signature. Three agents priced themselves in x402 anyway, using fields the schema does not define.