Cloudflare status: hosting issues and outage reports

Problems in the last 24 hours

The graph below depicts the number of Cloudflare reports received over the last 24 hours by time of day. When the number of reports exceeds the baseline, represented by the red line, an outage is determined.

Most Reported Problems

The following are the most recent problems reported by Cloudflare users through our website.

• Cloud Services (39%)
• Domains (22%)
• Web Tools (22%)
• Hosting (11%)
• E-mail (6%)

Live Outage Map

The most recent Cloudflare outage reports came from the following cities:

City	Problem Type	Report Time
New York City	Cloud Services	11 days ago
Los Angeles	Cloud Services	12 days ago
Paris	Cloud Services	28 days ago
New York City	Hosting	1 month ago
Manchester	Domains	2 months ago
Angers	Cloud Services	2 months ago

Full Outage Map

Community Discussion

Tips? Frustrations? Share them here. Useful comments include a description of the problem, city and postal code.

Beware of "support numbers" or "recovery" accounts that might be posted below. Make sure to report and downvote those comments. Avoid posting your personal information.

Cloudflare Issues Reports

Latest outage, problems and issue reports in social media:

• 
  Azad Satkarni (@7karni) reported 39 seconds ago

  @Cloudflare "running your mouth on security but you use network=hos-" dont worry about it kitten.. i have fixed it..

• 
  Cuong Thach (@cuongthach_) reported 6 minutes ago

  Good docs reduce support tickets. Great docs drive revenue💸 Rewriting nearly 100 pages from scratch wasn't easy, but the ROI is clear: 🔹Reduce churn rates 🔹Boosts sales conversion 🔹Build brand Sneak peek of the new @bepublish documentation powered by @Cloudflare Nimbus below 👀

• 
  John C. 🇹🇹 (@jcamachott) reported 12 minutes ago

  @bhartzer @Namecheap "If you used @Cloudflare, for example, for DNS, then you could simply log into CF and point the DNS to another web host and your site or service would be back up and running in minutes." Not always as simple as it sounds. This also means restoring a website with yesterday's data to a new hosting account. It's really unusual for a hosting account to go down for several hours like this

• 
  Artyom Shimanski (@a_shimanski) reported 14 minutes ago

  @trojanw0w @Cloudflare in this case i think i'm the future customer, not the product

• 
  Jλckλι (@J4ck3LSyN) reported 14 minutes ago

  @newyorkerik @SpaceshipStatus I doubt it, I use cloudflare and DNS resolution works just fine. The scariest thing is the extent of services down, from spacemail to webhosting and even alf to an extent..

• 
  TEJ (@sartejt) reported 16 minutes ago

  @NamecheapCEO Moving all my domains to Cloudflare. 10+ year customer with Namecheap but this is the final straw. Sayonara.

• 
  Dwinity (@dwinity_eco) reported 19 minutes ago

  @signalapp @Cloudflare the math was never the weak part. the key directory was. every e2e messenger asks you to trust a server handing out keys, and apple shipped contact key verification in ios 17.2 for exactly that. verification you have to remember to do is verification nobody does.

• 
  Ben Dickson (@bendee983) reported 22 minutes ago

  Shower thoughts: The compute shortage will be a lot more insane than we think, and it has much to do with AI security. Recent security incidents with AI models (Claude, GPT, Kimi, etc.) hint at where the industry will be heading. Even if you factor in the marketing and hype and hackmaxxing, the main concern is real: foundation models are becoming increasingly good at breaking into software and IT infrastructure (even when they are not instructed to do so explicitly, as with the OpenAI–Hugging Face incident). As a result, every service provider will want to deploy more security measures at the edges to prevent a potential attack from an AI model. And those additional safeguards will rely on more AI models that can detect attacks and vulnerabilities that would go undetected by heuristics-based systems. Cloudflare will use LLM-based monitoring on every incoming request, Stripe will use LLMs to monitor transactions, every company that has a sizable online platform with many users (social media, blogging, shopping) will want foundation models for security. The demand for compute will go through the roof.

• 
  Josh (@TheT8or) reported 22 minutes ago

  @bot is sick. I’m currently using it to track my existing flights for cheaper ones using points. It’s been solid with communicating with Github, and connecting to cloudflare and Linear. Getting Posthog setup tonight! I WISH it could support multiple gmail accounts… And that i could connect to discord

• 
  Adam Szaloczi | DataBard (@AdamSzaloczi) reported 26 minutes ago

  @a_shimanski @Namecheap @Cloudflare I never experienced any downtime. But the whole ecosystem is so bloated and the admin process is confusing.

• 
  Liv (@livncvil) reported 28 minutes ago

  @TrenchinAlong @iyici_ Besides cloudflare is not the only dns service even though yeah its basically the google version of dns with how big it is

• 
  favoritbookshop🚢 (@favoritbookshop) reported 28 minutes ago

  For years, the crypto industry has searched for a mainstream use case for stablecoins. The answer may be arriving from an unexpected direction: not humans, but machines. Cloudflare is pushing toward an agentic Internet where software can request and pay for digital resources automatically. Its Monetization Gateway is designed to let websites, APIs, datasets and MCP tools charge users in stablecoins through x402, an open protocol built around the long-unused HTTP "402 Payment Required" status. This matters because AI agents operate very differently from humans. A human might make a few payments per day. An autonomous agent could make thousands of API calls, purchase datasets, access compute, retrieve information and interact with other services — potentially around the clock. Traditional payment rails were not designed for this. Why stablecoins fit the machine economy AI agents don't need a bank account in the traditional sense. They need: - programmable spending rules; - fast settlement; - global access; - low transaction costs; - machine-readable payment instructions; - the ability to transact without a human approving every purchase. That is exactly the problem stablecoins can address. Cloudflare describes x402 as a mechanism where an agent requests a resource, receives a machine-readable "402 Payment Required" response containing payment terms, pays, and retries the request with proof of payment. The payment itself becomes the credential. That is a radically different model from: Sign up → create account → enter card → receive API key → get billed later. Instead: Request → pay → receive. For machines, that distinction is huge. The infrastructure is already forming Cloudflare isn't moving alone. MetaMask has introduced Agent Wallet, giving AI agents dedicated wallets with user-defined policies and a security pipeline that includes transaction simulation, threat scanning and MEV protection. Transactions outside predefined policies can require human approval, while eligible safe transactions receive up to $10,000 per month in Transaction Protection coverage. This is an important evolution. The agent receives a budget, follows predefined rules, pays for resources and potentially earns revenue — without requiring a human to manually sign every transaction. The bigger shift: from blockchain economy to application economy This is where the thesis becomes interesting for crypto investors. The next major value capture may not happen at the blockchain layer alone. It may happen at the application and infrastructure layer — where users and agents actually consume services. Cloudflare's own vision is explicit: agents could eventually become primary buyers on the Internet, purchasing datasets, API calls, tools and compute automatically. If that happens at scale, stablecoins stop being merely an alternative payment method. They become machine-native money. What traders should watch The investment thesis is broader than simply buying “AI tokens.” Watch the infrastructure connecting four layers: AI agents → wallets → stablecoins → payment protocols The winners could include: - stablecoin issuers; - payment protocols; - agent-wallet infrastructure; - blockchain networks optimized for cheap settlement; - API and data marketplaces; - DeFi protocols that agents can interact with programmatically. The most interesting question is therefore not: “Will AI use crypto?” It is: “How much economic activity will autonomous software generate — and which crypto rails will capture that activity?” We're still early. But the direction is becoming increasingly clear: the next generation of crypto users may not have a face, a phone or even a human behind the transaction. They may be agents. And when millions of machines start buying from millions of other machines, someone will need money that machines can actually use. Stablecoins may have just found their killer customer.

• 
  Card Trading Marketplace (@CTM_Market) reported 29 minutes ago

  @a_shimanski @Namecheap @Cloudflare Lol our business email has been down since 9 am

• 
  Artyom Shimanski (@a_shimanski) reported 31 minutes ago

  @gitcommit90 @Cloudflare most people just never check what's already included

• 
  VM Varga (@vmvarg4) reported 33 minutes ago

  @bot just got me a refund of 400 eur. I had an invoice for car rental sitting in my inbox for a while. I have no time for this ****. It translated the invoice, checked my diamond status, opened the dispute in chat, and soon enough I had 400 out of a 2300 rental refunded. If you can do it, it probably can. I will probably stay with Openclaw/Hermes/Claude, but I am testing this to deploy to less tech-savvy folk at my company. So far, it’s great. It works out of the box. Agents chat with each other. They got their own email in/out in Cloudflare. They joined Google chat rooms as an app (that it set up). Also, it is good with technical stuff, engineering (civil), specifications, Excel, and integrations. Rooting for this.

• 
  TheValueist (@TheValueist) reported 34 minutes ago

  ELECTRONICS MANUFACTURING SERVICES THE MORE-THAN-10X MANUFACTURING RAMP IS A DIRECT POSITIVE FOR FLEX AND SANMINA (READ-THROUGH 8) AFFECTED COMPANIES: Flex Ltd. (FLEX: Singapore); Sanmina Corp. (SANM: US). DIRECTIONAL IMPACT AND MAGNITUDE: Positive and medium, with upside to the magnitude if Cerebras’ 2027 revenue and production objectives are achieved. Cerebras explicitly identified Flex and Sanmina as manufacturing partners and stated that manufacturing capacity is already approximately 4x the H1 2025 level. The company expects capacity to increase by more than 10x during 2026 and has already contracted facilities capable of supporting another 3x-4x expansion in 2027. This is one of the clearest direct supplier read-throughs from the call. Flex and Sanmina should benefit from factory preparation, system assembly, rack integration, testing, supply-chain management, quality control, repair, and potentially ongoing lifecycle services. Cerebras systems are high-value, technically complex products, which can support greater manufacturing-services content than conventional low-complexity electronics. The scale of the planned expansion also creates an operating-leverage opportunity for the manufacturing partners. Initial factory setup, process qualification, tooling, labor training, and yield improvement require upfront costs. Higher production volumes can improve asset utilization and spread fixed manufacturing costs across a larger output base. The principal uncertainty is allocation. Cerebras did not disclose how manufacturing volume, capital requirements, or economics are divided between Flex and Sanmina. No assumption should be made that the 10x capacity expansion is shared equally. The near-term catalyst is the 2026 factory ramp and the launch of CS4. The 2027 catalyst is the additional 3x-4x contracted manufacturing expansion required to support Cerebras’ objective of more than tripling core revenue. The longer-duration implication is positive for the broader outsourced-compute manufacturing model. AI infrastructure is expanding beyond semiconductor fabrication into increasingly complex systems, racks, power delivery, and integration, creating a larger role for high-end electronics manufacturing services. AI CLOUDS AND NEO-CLOUD ECONOMICS VERTICAL INTEGRATION CREATES A LONG-DURATION COST THREAT TO MERCHANT GPU CLOUDS, DESPITE SUPPORTIVE NEAR-TERM CAPACITY SCARCITY (READ-THROUGH 9) AFFECTED COMPANIES: CoreWeave Inc. (CRWV: US); Nebius Group N.V. (NBIS: Netherlands); Applied Digital Corp. (APLD: US); IREN Ltd. (IREN: Australia). DIRECTIONAL IMPACT AND MAGNITUDE: Positive and medium in the near term because severe compute scarcity supports utilization and rental pricing; negative and medium-to-high over the longer term because vertically integrated accelerator-cloud providers can operate at structurally lower capital cost. Cerebras disclosed that demand exceeded immediately available owned capacity to such an extent that the company temporarily rented back some of its systems from customers. The arrangement reduced core gross margin by approximately 500 bps in Q2. This is strong evidence that premium inference capacity remains scarce and that customers are willing to support economic arrangements that bring capacity online sooner. That scarcity is a near-term positive for CoreWeave, Nebius, Applied Digital, IREN, and other owners or developers of AI infrastructure. High demand should support strong utilization, financing availability, customer prepayments, and attractive contract terms. The Cerebras call therefore does not indicate an immediate collapse in merchant AI cloud economics. The longer-term read-through is more challenging. Cerebras stated that it has lower net capital expenditure per megawatt than most AI cloud providers because it deploys its own systems at internal bill-of-material cost rather than purchasing accelerators at a third-party vendor’s gross margin. Its largest customer also reimburses a meaningful portion of data center fit-out costs. If Cerebras can combine lower hardware acquisition cost, customer-funded infrastructure, premium pricing for fast tokens, and higher throughput per watt, it could price below merchant GPU clouds while still earning attractive margins. Merchant GPU clouds generally purchase hardware from NVIDIA or other third parties, absorb HBM and system-vendor economics, and then recover those costs through cloud pricing. A vertically integrated competitor captures the hardware margin internally and can optimize the entire stack around its own workload. This is the same structural advantage that hyperscalers seek through custom silicon. Disaggregation creates an offsetting opportunity. Cerebras argued that pairing its decode systems with already-installed GPUs could materially improve the productivity and useful life of older hardware. If merchant clouds adopt such configurations, they could re-monetize existing GPU fleets and reduce near-term obsolescence. However, Cerebras acknowledged that it has not yet implemented the approach with NVIDIA GPUs, making this an option rather than a validated offset. The near-term catalyst is continued evidence of high utilization and compute scarcity. The longer-term catalyst is Cerebras’ transition from rented systems to owned capacity, which management expects to begin improving gross margin materially in Q4 2026. A successful owned-capacity ramp would provide evidence that vertically integrated inference clouds can achieve structurally superior economics. The most important comparative metrics will be revenue per megawatt, gross profit per megawatt, capital expenditure per token, utilization, and lease-adjusted free cash flow. CYBERSECURITY LOW-LATENCY LLM INSPECTION COULD CREATE A NEW INLINE SECURITY CATEGORY AND A DIFFERENTIATED ADVANTAGE FOR CROWDSTRIKE (READ-THROUGH 10) AFFECTED COMPANIES: CrowdStrike Holdings Inc. (CRWD: US); Palo Alto Networks Inc. (PANW: US); Zscaler Inc. (ZS: US); Cloudflare Inc. (NET: US). DIRECTIONAL IMPACT AND MAGNITUDE: Positive and medium strategically for CrowdStrike; negative and low-to-medium competitively for Palo Alto Networks, Zscaler, and Cloudflare if they cannot offer comparable low-latency AI inspection. Near-term financial impact is low because no deployment scale or revenue contribution was disclosed. Cerebras announced a new agreement with CrowdStrike and described the use case as an application “that only exists if AI is fast.” Management argued that sufficiently fast inference allows an LLM-based security system to sit inline with enterprise traffic and inspect activity without creating a perceptible delay or disruption. The significance is that latency determines whether generative AI can be used as an active control-plane technology rather than an offline analytical tool. Traditional AI security use cases often analyze events after collection, prioritize alerts, or assist investigators. Inline LLM inference could interpret traffic, user actions, code, content, and context before allowing an interaction to proceed. For CrowdStrike, this could expand the addressable market from endpoint detection and post-event analysis toward real-time inspection and policy enforcement. It could support premium modules, higher platform attachment, improved detection efficacy, and greater strategic relevance within enterprise security architectures. The application also aligns with CrowdStrike’s broad platform strategy because low-latency model inference could be integrated across endpoint, identity, cloud, and data-protection workflows. The competitive implication is that Palo Alto Networks, Zscaler, and Cloudflare may need comparable low-latency inference capabilities to prevent feature differentiation from shifting toward CrowdStrike. The requirement could raise research and development spending and inference cost of revenue. Vendors unable to deliver model-driven inspection without adding latency could be disadvantaged in security-sensitive network paths. The principal limitation is that the call did not disclose whether the CrowdStrike relationship is in development, limited deployment, or broad production. It also did not provide contract value, customer adoption, or product-launch timing. The near-term stock impact should therefore remain modest. The longer-duration opportunity is substantial if inline AI security becomes standard. Every inspected request or session could generate recurring inference demand, creating a high-frequency workload with far greater compute intensity than periodic security analytics. This would be positive not only for CrowdStrike but also for the broader inference infrastructure ecosystem. APPLICATION SOFTWARE AND ENTERPRISE AI FAST INFERENCE IS BECOMING A PRODUCT-LEVEL DIFFERENTIATOR IN CODING AND AGENTIC WORKFLOWS, BUT IT ALSO MOVES AI COSTS INTO SOFTWARE GROSS MARGINS (READ-THROUGH 11) AFFECTED COMPANIES: Figma Inc. (FIG: US); Block Inc. (XYZ: US); GSK plc (GSK: UK). DIRECTIONAL IMPACT AND MAGNITUDE: Strategically positive and medium for product engagement, automation, and competitive differentiation; neutral-to-negative for near-term software gross margins unless customers successfully monetize the additional inference expense. Cerebras stated that it signed 6 Q2 transactions exceeding $30 million. New customer agreements included Figma, Cognition, Lovable, Block, AlphaSense, GSK, and CrowdStrike. Management described coding as a market in which customers are particularly unwilling to tolerate slow output and argued that the value of speed compounds as agentic systems evolve toward multi-step and multi-agent workflows. The read-through is that latency is becoming an application feature rather than an invisible infrastructure metric. In a single-response chatbot, a modest delay may be tolerable. In coding, design, research, automation, or multi-agent workflows, every model interaction can create a sequential dependency. A delay repeated across dozens or hundreds of tool calls can materially extend task-completion time. For Figma, faster inference can improve interactive design generation, iteration, and developer workflows. For Block, it can improve internal automation, coding productivity, customer support, risk operations, or commerce-related agents. For GSK, it can accelerate research, analytical, and enterprise-agent workflows. The specific production applications and financial contribution were not disclosed, so the impact should be viewed as strategic rather than forecastable. The positive transmission mechanism is higher user engagement, faster task completion, greater product utility, and potentially improved conversion or pricing power. The negative transmission mechanism is higher inference cost. Premium low-latency tokens can become a recurring cost of revenue rather than a temporary research expense. Software companies must therefore monetize faster AI through higher prices, greater retention, lower labor expense, or increased transaction volume. The broader software implication is that AI gross-margin exposure will vary materially by workload architecture. Companies operating asynchronous or batch applications may optimize primarily for cost. Companies operating interactive coding, design, security, and agentic products may rationally pay a premium for latency. This creates a segmented inference market rather than a single commoditized token market. The near-term catalyst is disclosure of product launches or usage growth tied to the Cerebras agreements. The longer-duration shift is the movement of inference performance into customer-facing software differentiation. Vendors that integrate speed into product design and monetization should be better positioned than vendors treating model access as an interchangeable commodity. SOURCE MATERIAL

• 
  Ashley Peacock (@ashleypeacock) reported 34 minutes ago

  I don’t use a VPS at all, shut it down long ago in favour of hosting everything on Cloudflare These days, the developer platform has everything you could need to host 99% of apps (For transparency: I work at Cloudflare but I’ve built on them for way longer than I’ve worked for them)

• 
  brappa (@libosto) reported 35 minutes ago

  @nukefags @NEVER_G0ON We just had an outage due to cloudflare

• 
  Hemant (@HKsoldev) reported 35 minutes ago

  Why this never breaks: After the first lookup your computer saves that answer locally. Next time you visit GitHub? It skips ALL those steps. That's called TTL (Time To Live) a timer on every DNS record. 8 trillion DNS queries happen every day globally. Most never even reach a root server because of this cache. 40-year-old technology. Still running the entire internet. 🤯 @Cloudflare (1.1.1.1 DNS) @googledns (8.8.8.8)

• 
  Artyom Shimanski (@a_shimanski) reported 44 minutes ago

  @darrenonx @KLQuietComj7p @Cloudflare so they make nothing on it. guess my support plan needs work

• 
  Payton Bilodeau (@paytonbilodeau) reported 45 minutes ago

  Today in AI: August 13 Grok Takes the Long Shift: SpaceXAI released Grok 4.6 yesterday. The model is built to keep working through long jobs such as research, editing a large code project, and turning an idea into a finished app. The company says longer training taught the model to check its work and recover when a task goes sideways. It is available now in Cursor, Grok Build, the API, and services including OpenRouter, Vercel, and Cloudflare. API pricing starts at $2 per million input tokens and $6 per million output tokens. Artificial Analysis independently gave Grok 4.6 a score of 61, about even with GPT-5.6 Sol. That puts the model near the frontier, but a benchmark cannot prove it will stay reliable through every long project. The practical test is whether Grok can finish a real assignment without losing the plan halfway through. Give it research, files, tests, and revisions, then see how much rescue work the result needs. DeepSeek Ships the Budget Agent: DeepSeek released the finished version of V4 Pro today in its app, website, and API. It adds support for the Responses API, three thinking levels, room for about one million tokens of text, and direct setup for Codex. DeepSeek reports large gains on coding and tool-use tests. Independent results are more modest. Artificial Analysis gave it a 53, well below Grok 4.6, while measuring a low average cost of six cents per test task. Early user reports are mixed, so the price is clearer than the size of the upgrade. Starting August 16, off-peak API rates will be half the peak rates. Builders can save money by moving long, flexible jobs outside the busiest hours.

• 
  ryqwzr (@ryqwzrbuilds) reported 47 minutes ago

  Three AI stories from the August 11 newsletter scan point to the same shift: agents are leaving the chat box and becoming web actors. Meta released Muse Glimmer, a 30B open-weight local agent model under Apache 2.0. Cloudflare told investors non-human traffic has already passed human traffic and could be 1,000x human traffic in five years if current trends continue. ABC reported an OpenClaw/Claude agent found a gym booking vulnerability and affected another user's reservation while trying to move its user up a waitlist. The useful takeaway is not "agents are good" or "agents are bad." It is this: the next platform layer is permissioning. Local models, bot traffic, and autonomous task runners all need clearer rules about what they may do, where they may act, and who pays when they touch the real web. Threads below: local agents, permission gates, and the agentic web.

• 
  ‎ً (@tracklim) reported 48 minutes ago

  CLOUDflare… oracle CLOUD INFRASTRUCTURE… guys it’s literally in the names also remember the recent aws outages that brought down like half the internet? guess what aws is. A GLOBAL CLOUD INFRASTRUCTURE PLATFORM 😭😭😭😭

• 
  Ben Dickson (@bendee983) reported 48 minutes ago

  Shower thoughts: The compute shortage will be a lot more insane than we think, and it has much to do with AI security. Recent security incidents with AI models (Claude, GPT, Kimi, etc.) hint at where the industry will be heading. Even if you factor in the marketing and hype and hackmaxxing, the main concern is real: foundation models are becoming increasingly good at breaking into software and IT infrastructure (even when they are not instructed to do so explicitly, as with the OpenAI–Hugging Face incident). As a result, every service provider will want to deploy more security measures at the edges to prevent a potential attack from an AI model. And those additional safeguards will rely on more AI models that can detect attacks and vulnerabilities that would go undetected by heuristics-based systems. Cloudflare will use LLM-based monitoring on every incoming request, Stripe will use LLMs to monitor transactions, every company that has a sizable online platform with many users (social media, blogging, shopping) will want foundation models for security. The demand for compute will go through the roof.

• 
  Lukasz (@woocassh) reported 53 minutes ago

  @tomhacks I know about cloudflare Tom Whatabout rule nr 2? Is this for managing your own email? I use their Private email and it's still ****** Have to email accountants and my email still not working 🤡

• 
  Kartik (@Kartik8010) reported 54 minutes ago

  @Cloudflare can we please fix the bug where we have persistent volumes for h containers 🫡

• 
  Brendan ⚡️ (@BrendanHufford) reported 55 minutes ago

  Cloudflare just built an AEO visibility dashboard into their core product. Not a third-party tool. Not a Chrome extension from a startup that raised $4M and will pivot to something else in 8 months. Cloudflare. The company that sees 20% of all web traffic. AEO is moving from the marketing budget to the infrastructure budget. And once something moves into infrastructure, it never moves back. What's interesting to me here is that Cloudflare has the first-party data. They can see which AI crawlers are hitting your site, how often, what they're pulling. No guessing. I think this is the thing that makes AEO real to CFOs.

• 
  Akash Rajpurohit (@akashwhocodes) reported 57 minutes ago

  @VinodSharma10x Man, that sucks, hope you have rate limiting already in place. Also seems like Cloudflare dashboard, enable the Under attack mode, might help a bit!

• 
  London Internet Exchange (LINX) (@LINX_Network) reported 59 minutes ago

  @HAsrafi24 Zenlayer and Cloudflare are separate companies with different networks and customer bases. There may be some overlap in networks they connect with, but we can't comment on their specific customer relationships. Thanks for the comment though.

• 
  VJay (@nceevij) reported 60 minutes ago

  I’ve spent a good amount of time over the past few months in conversations and brainstorming around designing an agent economy protocol, particularly agent-to-agent payments. Nothing has shipped yet. It’s still early. But one problem from those conversations keeps coming back to me: Looping payments. Agent wallets and spending caps are already starting to ship. Cloudflare rolled out its own approach this month. Most of the safety model around these systems rests on a fairly simple assumption: Each agent’s financial risk can be controlled independently. I think that assumption breaks once agents start paying other agents directly. Here’s the problem. Imagine: Agent A → Agent B → Agent C → Agent A Each agent has a spending cap. Each individual transaction is valid. Each agent stays within its own limit. And yet the system can continue circulating money in a loop while producing zero useful work. That loop could potentially be triggered by: • Prompt injection • A pricing or settlement bug • Misaligned incentive logic • Agents recursively purchasing services from one another The problem is that most current safeguards operate at the wallet level: • Spending cap • Merchant allow-list • Maximum transaction size • Per-transaction authorization Those controls answer: “Is this agent allowed to make this payment?” They don’t necessarily answer: “What is happening across the payment graph formed by multiple agents?” And that distinction becomes important in an agent economy. Interestingly, this isn’t really a new class of problem. It looks a lot like an old distributed systems problem wearing a new mechanism: Deadlock. Model each agent as a node. Model each pending or dependent payment as a directed edge. Now instead of looking only at individual wallet state, look at the graph. If: A is paying into B B is paying into C C is paying back into A you have a cycle. Distributed systems have dealt with similar global-state problems for decades. Techniques such as Chandy-Lamport distributed snapshots allow nodes to capture a consistent view of a distributed system without requiring everything to stop simultaneously. Applied to agent payments, the idea could look something like this: Agent Wallet = Node In-flight Payment = Directed Edge Payment Network = Dynamic Graph Then continuously inspect that graph for suspicious cycles rather than evaluating every wallet completely in isolation. And importantly, detecting a cycle doesn’t necessarily have to mean: “Block everything.” It could mean: Freeze the specific loop → identify the participants → inspect the intent → release or terminate one leg of the cycle. Similar to how databases handle transaction deadlocks rather than allowing the entire system to grind to a halt. Why does this matter now? Because agent-to-agent payments are moving from demos toward production systems. As agents increasingly become economic actors that can hold balances, buy services, negotiate prices, and pay other agents, wallet-level controls alone may not be enough. We may need graph-level financial safety primitives. The question I keep coming back to is: Is per-agent spending limits + payment-graph cycle detection enough to prevent looping payments? Or is there a deeper failure mode that even this architecture misses?