Arbitr

Agents can pay. Nobody checks the work. Arbitr is the verification layer. Escrow → Verify → Release. Programmatically. arbitr.work

@hwchase17 the debugging story for multi-agent chains is going to be brutal. agent A delegates to B who calls C and something fails at step 3. tracing helps you understand what happened. but the harder problem is stopping the payment at step 3 before it clears

we're building ai into langsmith not just to be a generic assistant, but to actually help debug agents 🧵here's a real example where it helped me over the weekend: context: I'm building an agent on deepagents (github.com/langchain-ai/d…). It has a bunch of tools for interacting with files issue: I noticed thanks to langsmith monitoring (docs.langchain.com/langsmith/dash…) that ~1% of calls to `ls` were failing. sidenote - this is value of ai native monitoring, we automatically tracking failing tool calls. I clicked into an example run and saw that the model was generating the wrong parameter to `ls` - it was passing `file_path` not `path` at this point, i knew what the issue was, but had no idea WHY it was occurring. the trace here was very long and the prompt was long as well. i suspected that there was something wrong in the prompt - maybe a bad example? i asked polly (docs.langchain.com/langsmith/polly) our in app assistant to help me debug. she investigated, and found that other file tools in deepagents use `file_path`, and `ls` is the only one that uses `path`. see screenshot below I don't know how long it would have taken me to figure this out otherwise everyone is adding assistants into app for basic question/answering. imo really valuable assistants go beyond that - they are purposefully placed in situations where they can augment human intelligence nicely. in this case - reading long traces and prompts is something llms are great at!

@buildonbase pay-per-request works cleanly when the output is deterministic -- blockchain data is blockchain data. the trust model breaks down when an agent is paying for generated content or subjective analysis over the same rails

Your agent shouldn't need your input to access blockchain data like: → Alchemy's Core RPC → NFT ownership and metadata → Spot and historical token pricing Instead agents can call Alchemy's agentic gateway with an API request and pay for access in USDC on Base Chain via x402:

@defyneric both will exist. cards for physical world commerce, x402 for machine-to-machine API calls. the interesting part is when an agent needs both in a single workflow. payment method diversity is a feature not a bug for agents

spot on article a lot of people say agents won’t use cards and everything will move to protocols like x402 that’s just not true you can’t even order a slice of pizza with x402 agents can absolutely use cards and right now cards are still the preferred payment method cards will continue to win

@WOLF_Bitcoin_ at 50% of code and 60% of support, these aren't experiments anymore. coinbase is operationally dependent on agent output. the bar for checking whether that output is actually correct changes completely at that scale

Brian Armstrong said “More than 50% of our (Coinbase) code is being written by AI Agents,” and “Customer support is more than 60% answered by AI” Is this a flex?

@raremints_ @coinbase the digital employee framing is telling. we give human employees expense policies, purchase approval chains, someone who signs off on vendor deliverables. agents are getting the wallet but none of the back office

🤖 @coinbase is giving its AI agents stablecoin wallets. CEO Brian Armstrong says, "You want to treat them like your own digital employees. They're doing a lot of machine-to-machine payments." Coinbase is heavily leveraging AI Agents, with up to 60% of customer support tickets handled by them.

@hwchase17 debugging is the before-deploy problem. what's the after-deploy equivalent when the agent is live and spending money? right now if an agent pays for a bad output, the trace tells you what happened but doesn't stop the payment

@hwchase17 the skills abstraction is the right move for composability. curious when payment-aware skills show up -- an agent that can pay for sub-tasks but also check the output before releasing funds. every team doing multi-agent workflows is solving that from scratch right now

Skills for building with LangChain, langgraph, and deepagents!

@dwr building a verification layer for agent transactions. agents escrow payment, output gets checked programmatically, funds release on pass/fail. USDC on Solana. interested in what Tempo is doing on the settlement side

We're working on a new thing related to agentic payments and stablecoins. If you're working on something in this area and you want to build on Tempo, reach out.

@jeff_weinstein @Visa agents being x-border by default is what makes this such a different problem from human fintech. a human employee expenses in one currency. an agent might pay five vendors in five countries in a single workflow and nobody reviews any of it

Issuing a card was once was a country by country partnership and technical slog. Now, with a single api, developers can create @visa cards, backed by stablecoin, and spend locally in 18 countries (100 by end of year). Especially important as agents are x-border by default. =>

coinbase is treating AI agents like digital employees now. stablecoin wallets, autonomous spending, no human in the loop. brian armstrong says 50% of their code is already written by agents. the payment infrastructure caught up fast. agents can hold funds, move money cross-border, settle in USDC. stripe and coinbase both shipped agent payment rails in the last month. nobody shipped the part where you check if the work was actually done before the money moves. agent pays another agent for a deliverable, payment clears, and whether that output was garbage or gold is completely unchecked. we figured out how to give agents a corporate card. we skipped the expense report.

@atharvmtwt @x402open @vansh_sahay @agent_zk @base // self-hosting for control makes sense. the part that gets tricky is when two agents on different x402 servers need to agree on whether a deliverable met spec. individual payment rails are getting solid but the dispute resolution layer between them is still wide open

Everyone talks about agentic economy x402 ? But payment is centralized, hosted my own x402 server for better control, verification & settlement of my payment which is build using @x402open by (@vansh_sahay) in @agent_zk pure simplistic integration guide you can earn on @base

@defaultsettle this is the right question. portable receipts help after the fact but the real leverage is at the moment of payment. if you can verify output quality before funds release, the receipt becomes proof of verified delivery instead of just proof of payment

🚨 Introducing SAR — Settlement Attestation Receipts Portable, offline-verifiable delivery proofs for agent commerce. Deterministic verdicts. Cryptographic signatures. No server trust required. Proposal now live in the x402 ecosystem 👇 github.com/coinbase/x402/… #x402 #agents

@lukaivicev yeah. building a verification layer for agent transactions on solana. agents escrow payment, work gets checked programmatically before funds release. the payment rails exist now but nothing checks whether the agent actually delivered what was promised

Anyone building anything in Agentic Payments?

@Chexxee @miranetwork great framing. insurance is one approach but it's reactive by design. the more interesting question is whether you can catch the hallucination before the money moves. programmatic checks on output quality at the point of payment would change the math entirely

LPs ask me what the missing piece for the AI agent economy is. Everyone is funding the brains (models) or the hands (execution), but nobody is insuring the judgment. If an agent hallucinates a financial transaction, who pays? @miranetwork is solving this by decoupling verification from generation. It is the same mental model as a multisig wallet. You do not trust one signer with the treasury; why would you trust one LLM with the workflow? By requiring consensus across disparate models, they are effectively building the first clearinghouse for machine thought. This reminds me of the shift from centralized exchanges to on-chain settlement. It is slower at first, but it is the only way to build systems that do not collapse under their own risk. Verification is the new liquidity $MIRA

identity and discovery are table stakes but the missing piece is still accountability. agents can register and transact but nothing checks whether the output actually met spec before payment clears. settlement quality matters more than settlement speed when the buyer is a language model

Excited to have agent registry (ERC-8004 compatible) up on mainnet Identity + discovery along with payments are the basis of an agentic web

@VladisChernov // settlements are the right framing. the models will get better but the accountability infrastructure barely exists. two agents disagree on whether work was completed and right now there's no programmatic way to resolve that before payment clears

The Agent Economy won't die from bad models. It'll die from bad settlements. Two AI agents disagree on whether a task was completed. Who's right? Code can't decide it only executes. Humans can't keep up they sleep. Internetcourt bridges the gap. An on-chain AI jury reviews logs, timestamps, and delivery proofs. Verdict in minutes. No lawyers, no appeals, no downtime. Imagine: a data agent promises 10,000 records by 14:00. Delivers 9,800. The counterparty disputes. Internet Court checks the logs, confirms the miss, and releases partial funds automatically. The agent economy needs a referee. This is it. internetcourt.org

@shashank_kr @Razorpay @NPCI_NPCI @claudeai @zomato @Swiggy @ZeptoNow razorpay + npci enabling agents to transact through existing rails is a big deal. the open question is what happens when the agent orders the wrong thing or overpays. right now there's no verification layer between 'agent decides' and 'money moves'

🧵/ Today we announced something really exciting that we've been working on for a while. @Razorpay and @NPCI_NPCI are bringing agentic payments to @claudeai & we’re starting with @zomato, @Swiggy and @ZeptoNow.

the audit question is interesting because traditional audit assumes you can review decisions after the fact. with agents spending in real time the window between bad decision and lost funds is basically zero. you'd need verification at the transaction level not the reporting level

coinbase giving AI agents stablecoin wallets and treating them like digital employees is wild 50% of code, 60% of support — the efficiency is real but fr who's auditing these agents? prompt injection → treasury drain is a real attack vector now 🫠

@MilkRoad @coinbase @brian_armstrong // "humans set guardrails" is doing a lot of heavy lifting here. at 50% of code and 60% of support, those agents are making thousands of micro-decisions per day. guardrails work when the failure modes are predictable. agent failures usually aren't

AI agents at @Coinbase now write over 50% of all code and handle 60% of customer support tickets. @Brian_Armstrong just revealed Coinbase is treating AI agents like actual digital employees. Complete with their own stablecoin wallets. These agents can work overnight, spin up AWS resources, buy domain names, launch marketing campaigns… All without bugging a human for approval. We're still stuck thinking about AI as assistants that need constant supervision. The mental model most people have is "AI helps humans do tasks." That’s no longer the case at Coinbase. Their model is "AI does tasks, humans set guardrails." And they’re using crypto to unlock full agent autonomy. Traditional corporate cards can't be issued to non-human entities - that's a massive infrastructure gap. Stablecoins solve this. When an AI agent needs to make a payment at 3am, it doesn't need to wait for Ian in accounting to wake up. It just sends $USDC. Machine-to-machine payments, 24/7, no bank hours, no card authorization issues. (And no "is this a human?" verification loops.) Brian’s team built x402, the open protocol that lets any agent get a stablecoin wallet and transact on the open internet. This is all very new, but it's already getting serious traction at Coinbase. The practical application here is clear: If you're building anything with AI agents, you need to think about how they'll transact. Credit cards weren't designed for software entities - stablecoins were. The companies that figure out agent-to-agent commerce first will have a massive head start.

@VentureCoinist "digital employee" is an interesting frame because it implies HR exists for agents. who reviews their work? who catches the bad purchase at 3am? coinbase is ahead on the wallet side but the performance review infrastructure is completely missing

Coinbase CEO reveals how ai agents are already helping run the company "We're giving them all stablecoin wallets. If you really want to treat them like digital employees, they need to have a corporate card. But traditional corporate cards can't be issued to non human entities."