<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Nimesh Tripathi on Medium]]></title>
        <description><![CDATA[Stories by Nimesh Tripathi on Medium]]></description>
        <link>https://medium.com/@nimeshtripthi?source=rss-40c2b309dabd------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*ZP_EEJ-iF695ZZajFFxOzA.png</url>
            <title>Stories by Nimesh Tripathi on Medium</title>
            <link>https://medium.com/@nimeshtripthi?source=rss-40c2b309dabd------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 13:01:13 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@nimeshtripthi/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Your Customers’ AI Agents Are Already Logging In. Now What?]]></title>
            <link>https://medium.com/@nimeshtripthi/your-customers-ai-agents-are-already-logging-in-now-what-846b0591bba6?source=rss-40c2b309dabd------2</link>
            <guid isPermaLink="false">https://medium.com/p/846b0591bba6</guid>
            <dc:creator><![CDATA[Nimesh Tripathi]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 01:33:42 GMT</pubDate>
            <atom:updated>2026-10-27T01:47:30.421Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Big banks are blocking AI agents. That’s a reasonable first move and a bad strategy.</em></p><p>Last week, an <a href="https://paymentexpert.com/2026/10/23/ai-agents-blocked-by-banks/">Axios report</a> said JPMorgan Chase, Bank of America, Citi and Morgan Stanley are restricting third-party AI agents from customer accounts. The trigger: agents logging in with customers’ own credentials, sometimes with the customer’s blessing, but in ways that break the bank’s terms of service.</p><p>We’ve seen this movie before. It was called screen scraping, and it took the industry roughly a decade, a standards body and a federal rulemaking to clean it up.</p><p>I’m a product manager in fintech, and my read is simple: blocking agents is a security response to what is actually a product problem. The customers who sent those agents didn’t do anything wrong. They just got there before we did.</p><h3>The real problem isn’t the agent. It’s that we can’t see it.</h3><p>According to the reporting, banks’ core complaint is that they can’t distinguish genuine, customer-approved agent activity from rogue activity. That’s the whole ballgame.</p><p>Today, when an agent uses a customer’s username and password, our systems see one thing: the customer. Same credentials, same session, same permissions. The bank has no way to know that:</p><p>• a piece of software, not a person, is acting</p><p>• which software it is, and who built it</p><p>• what the customer actually asked it to do</p><p>• whether the customer is still okay with it an hour later</p><p>Every fraud model, every step-up challenge and every dispute process we’ve built assumes a human on the other end. Credential-sharing agents quietly break all of those assumptions at once. So blocking them isn’t paranoia. It’s the only control available when you’re navigating a dark room.</p><p>But “we can’t see it” is a design flaw, and design flaws are fixable.</p><h3>Block vs. build is a product decision</h3><p>The instinct is to hand this to security and fraud teams. They’ll do what they’re paid to do: minimize risk. Blocking minimizes risk. But it doesn’t minimize demand.</p><p>Customers who want an agent to pay their bills, move money between accounts or dispute a charge won’t stop wanting that. They’ll do one of three things:</p><p>1. Keep sharing credentials and find workarounds, so the risk goes underground.</p><p>2. Move their primary relationship to a provider that welcomes agents.</p><p>3. Give up, and quietly decide your bank is the hard one to work with.</p><p>None of those are good outcomes. And the industry is already moving. Payments infrastructure players are publishing <a href="https://www.imf.org/-/media/files/publications/imf-notes/2026/english/insea2026004.pdf">agent payment standards</a>, and some banks are already putting <a href="https://www.pymnts.com/news/artificial-intelligence/2026/banks-add-ai-agents-keep-their-40-year-old-cores/">AI agents to work on their own systems</a>.</p><p>So the question for product leaders isn’t “should we allow agents?” It’s “what would it take to say yes safely?” The answer comes down to three strategic decisions: identity, permissions and consent.</p><h3>1. Agent identity: know who’s knocking</h3><p>The first fix is to stop letting agents pretend to be people. An agent should arrive with its own identity, separate from the customer’s, and that identity should answer three questions:</p><p>• <strong>Who built it?</strong> A registered developer or company you can hold accountable.</p><p>• <strong>Who is it acting for?</strong> A specific customer, bound to a specific consent.</p><p>• <strong>What is it?</strong> A known agent type with a declared purpose, not an anonymous script.</p><p>This is exactly the lesson of open banking. Credential sharing was replaced by tokenized, API-based access, where the bank always knew which third party was calling. Agents need the same treatment, with one twist: the “third party” may be a general-purpose assistant acting across dozens of tasks, not a single-purpose budgeting app.</p><p>For PMs, this means agent identity belongs in your auth roadmap, not in a fraud backlog. If your platform can’t issue, recognize and revoke an agent credential, every other control in this article is moot.</p><h3>2. Delegated permissions: scoped, revocable, progressive</h3><p>A customer’s password is an all-access pass. An agent should never get one. Instead, think of permissions as a ladder the agent climbs, one rung at a time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*chUY865rqCfETntOxZl5qg.png" /></figure><p>The banks that are already deploying agents internally show this works. One lender CEO quoted by <a href="https://www.pymnts.com/news/artificial-intelligence/2026/banks-add-ai-agents-keep-their-40-year-old-cores/">PYMNTS</a> described starting agents on narrow, repetitive tasks and widening access only as performance earns it. In their example, a collections agent can process a payment but cannot reverse one.</p><p>If banks trust that model for their own agents, they can offer a version of it to their customers’ agents. Three design principles carry most of the weight:</p><p>• <strong>Least privilege by default.</strong> Grant the narrowest scope that completes the task.</p><p>• <strong>Limits live on the mandate, not the account.</strong> “Up to $200 to my electric company, monthly” is a mandate. Account access is not.</p><ul><li><strong>Revocation is instant and visible.</strong> One tap kills a mandate, and the agent knows it immediately.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/1*XAF3SVSMUME5yAgLD1sN1g.png" /></figure><h3>3. Consent UX: mandates people actually understand</h3><p>Here’s where product teams earn their keep. Identity and permissions are plumbing. Consent is the moment the customer decides whether to trust it.</p><p>Most consent screens in finance are legal documents pretending to be UI. That won’t survive agents, because the stakes are higher and the actions are less predictable. A few patterns I’d design toward:</p><p>• <strong>Plain-language mandates.</strong> “Pay your Con Ed bill each month, up to $250, until you turn it off.” Not “Grant write access to payment initiation scopes.”</p><p>• <strong>Consent at the moment of intent.</strong> Ask when the agent first needs a capability, not in a 14-permission wall at setup.</p><p>• <strong>A single agent dashboard.</strong> Every connected agent, what it can do, what it did this week, and an off switch.</p><p>• <strong>Receipts the human can read.</strong> Each agent action logged in the customer’s activity feed, labeled as agent-initiated.</p><p>• <strong>Graceful step-up.</strong> When an agent hits a limit, it pings the customer for approval instead of failing silently or retrying.</p><p>There’s a liability angle here too. When something goes wrong, the dispute will hinge on what the customer authorized. A clear, logged mandate protects the customer, the bank and the agent provider. A shared password protects no one.</p><h3>What I’d ship in the next two quarters</h3><p>If I owned this roadmap today, I wouldn’t try to boil the ocean. I’d sequence it:</p><p>1. <strong>Detect.</strong> Instrument logins to flag likely agent sessions. You can’t manage what you can’t measure, and you’ll learn which tasks customers are delegating.</p><p>2. <strong>Talk to those customers.</strong> The people already using agents are your best research panel. Find out what jobs they’re hiring agents for.</p><p>3. <strong>Open a read-only agent lane.</strong> Registered agents, tokenized access, no money movement. Low risk, high learning.</p><p>4. <strong>Pilot one bounded mandate.</strong> Pick a single high-volume, low-risk job, like paying a known biller, and build the full consent loop around it.</p><p>5. <strong>Publish your agent policy.</strong> Tell developers how to get access legitimately. A clear front door is the best deterrent to people climbing through the window.</p><h3>The window is short</h3><p>Blocking buys time. It doesn’t buy a strategy. The banks that treat this as a product opportunity will define how agents bank. The rest will get standards and expectations handed to them by someone else.</p><p>Your customers’ agents are already logging in. The only question is whether they’re coming through a door you designed.</p><p><em>What’s your team doing about agent access? I’d love to hear how other PMs are approaching it.</em></p><h3>Sources</h3><p>• <a href="https://paymentexpert.com/2026/10/23/ai-agents-blocked-by-banks/">Payment Expert: AI agents blocked by JPMorgan, Citi and Bank of America</a></p><p>• <a href="https://www.pymnts.com/news/artificial-intelligence/2026/banks-add-ai-agents-keep-their-40-year-old-cores/">PYMNTS: Banks add AI agents, keep their 40-year-old cores</a></p><p>• <a href="https://www.imf.org/-/media/files/publications/imf-notes/2026/english/insea2026004.pdf">IMF Note: How agentic AI will reshape payments</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=846b0591bba6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Grocery Store Test: Explaining MCP, REST and JSON-RPC to Executives]]></title>
            <link>https://medium.com/@nimeshtripthi/the-grocery-store-test-explaining-mcp-rest-and-json-rpc-to-executives-eba9e4cd2d8b?source=rss-40c2b309dabd------2</link>
            <guid isPermaLink="false">https://medium.com/p/eba9e4cd2d8b</guid>
            <category><![CDATA[mcp-protocol]]></category>
            <category><![CDATA[rest-api]]></category>
            <category><![CDATA[gen-ai-for-business]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[solutions]]></category>
            <dc:creator><![CDATA[Nimesh Tripathi]]></dc:creator>
            <pubDate>Mon, 19 Oct 2026 17:09:54 GMT</pubDate>
            <atom:updated>2026-10-19T17:09:54.281Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How one executive’s question changed the way I explain technology, and the grocery store analogy that came out of it.</em></p><h3>The question</h3><p>As a product manager, I field a lot of questions. One from a senior business leader stayed with me long after the meeting ended.</p><p>He came from finance and operations: sharp, numbers-driven, and not technical by background. We were walking through our AI roadmap when he asked, “Everyone keeps saying MCP. We already have REST. Someone mentioned JSON-RPC. Why do we need another one?”</p><p>I gave him the answer I’d give an engineer:</p><blockquote>“REST is a resource-oriented style over HTTP, where you use verbs like GET and POST on endpoints like /invoices/123. JSON-RPC sends a payload like {“method”: “getInvoice”, “params”: {“id”: 123}} to a single endpoint. And MCP is built on JSON-RPC 2.0, adding a standard way for AI clients to discover tools and data exposed by servers…”</blockquote><p>It was accurate. He listened, nodded, and we moved on to the next agenda item.</p><h3>The realization</h3><p>On the way home, I kept replaying that moment. Somewhere around my second sentence, I’d watched his attention drift, and I realized why. He hadn’t asked <em>how</em> these protocols work. He’d asked <em>why</em> the business should care.</p><p>Leaders from finance and operations evaluate technology through cost, speed, risk and leverage. Words like “endpoint” and “payload” don’t map onto any of those. The gap in that room wasn’t his knowledge. It was my translation.</p><p>So I went looking for an explanation simple enough, and useful enough to win executive buy-in. I found it in the most ordinary place I could think of: the grocery store.</p><h3>The grocery store analogy</h3><p>Everyone has bought groceries. That’s what makes this work: nobody needs a glossary to follow along.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/864/1*dMabITXbnjtPDNf6RwzbOQ.png" /></figure><p>The grocery store analogy · REST, JSON-RPC and MCP</p><h3>REST: walk the aisles</h3><p>Every item lives at a known spot: Aisle 4, Dairy, milk. You go to that spot and do one of a few simple things. Look at it (GET). Put a new one in your cart (POST). Swap it for another (PUT). Put it back (DELETE). Where you’re standing tells the store what you’re dealing with.</p><p><strong>The business translation:</strong> REST is how most business software already talks. Your ERP, your CRM, your payment provider and your bank feeds almost certainly expose REST APIs. It’s the reliable, well-understood workhorse. If you’ve ever approved a “Salesforce integration”, you funded REST work.</p><h3>JSON-RPC: hand your list to the service counter</h3><p>Instead of walking the aisles, you go to one counter and hand over a neat slip: “Please do <em>fetchItem</em>: milk, 2 litres.” The clerk performs the action named on the slip and hands back a neat receipt. JSON is simply the tidy, labelled handwriting that both sides can read.</p><p><strong>The business translation:</strong> JSON-RPC is a lightweight messaging format, not a product. You rarely budget for it directly. It matters because it’s simple and predictable, which is exactly why MCP chose it as its language.</p><h3>MCP: one standard order window at every store</h3><p>Now picture a personal shopper, your AI assistant, who needs to shop at many stores: the bakery (your calendar), the pharmacy (your HR system), the hardware store (your finance system), the deli (your ticketing tool). Without a standard, the shopper has to learn each store’s unique ordering process. Someone has to set up, and keep maintaining, a custom arrangement with every single store.</p><p>MCP is like every store installing the same standard order window. Each window has a menu posted showing exactly what that store offers, so the shopper can walk up to any store and know how to order. The slips passed through that window are written in JSON-RPC. And behind the window, the store’s staff still walk the aisles, often using REST.</p><p><strong>The business translation:</strong> these three aren’t competitors. MCP <em>uses</em> JSON-RPC, and it usually sits in front of the REST APIs you already have. It isn’t replacing anything. It’s the layer that lets AI use everything you’ve already built.</p><h3>Why it matters: the integration math</h3><p>The grocery store gets a leader nodding. The math gets them to sign off.</p><p>Say your company uses 5 AI tools: a chat assistant, a coding assistant, a finance copilot, a support bot and an analytics agent. Say each needs access to 20 business systems: ERP, CRM, HRIS, data warehouse, ticketing and so on.</p><p>In grocery terms: that’s 5 personal shoppers and 20 stores. Without a standard order window, every shopper needs its own arrangement with every store.</p><p>Without a standard, every AI tool needs its own custom connector to every system. That’s 5 × 20 = <strong>100 connectors</strong> to build, secure, test and maintain. With MCP, each AI tool learns the plug once and each system exposes the plug once. That’s 5 + 20 = <strong>25</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/864/1*VzrJDIXYfRFybEuEUpQJXw.png" /></figure><p>Illustrative calculation · 5 AI tools, 5 to 40 business systems</p><p>The gap widens as you grow. Add a sixth AI tool tomorrow, and the custom path costs 20 new connectors. With MCP it costs one.</p><p>For a finance mind, that’s the pitch: <strong>MCP converts integration cost from multiplicative to additive.</strong> Fewer connectors means less engineering spend, faster rollout of new AI tools and a smaller surface to secure and audit. It also reduces vendor lock-in, because swapping one AI tool for another doesn’t mean rebuilding every connection.</p><h3>The real numbers</h3><p>Analogies open the door. Numbers walk the executive through it. Here’s what the data says as of October 2026.</p><p><strong>REST is the established default.</strong> Postman’s State of the API surveys have shown REST as the dominant API style for years: <a href="https://devops.com/2020-postman-state-of-the-api-reveals-api-investments-will-remain-strong/">93.4% of respondents used it in 2020</a>, and <a href="https://www.infoworld.com/article/2338668/ai-is-catching-on-with-api-developers.html">86% in 2023</a> as GraphQL and others grew. Your existing systems speak REST, and that isn’t changing.</p><p><strong>MCP went from idea to industry standard in about a year.</strong> Anthropic introduced MCP in November 2024. By December 2025, the project reported <a href="https://blog.modelcontextprotocol.io/posts/2025-12-09-mcp-joins-agentic-ai-foundation/">more than 97 million monthly SDK downloads and 10,000 active servers</a>, with support in ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot and Visual Studio Code.</p><p><strong>It’s now vendor-neutral.</strong> In December 2025, Anthropic <a href="https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation">donated MCP to the Agentic AI Foundation</a> under the Linux Foundation. The foundation was co-founded by Anthropic, Block and OpenAI, with support from Google, Microsoft, AWS, Cloudflare and Bloomberg. For a risk-minded executive, this matters: competitors jointly backing one standard means you’re not betting on a single vendor.</p><p><strong>Enterprises are moving past pilots.</strong> A Stacklok survey of senior technical leaders, <a href="https://www.digitalapplied.com/blog/mcp-adoption-statistics-2026-model-context-protocol">summarized by Digital Applied</a>, found most software organizations are already piloting or running MCP.</p><p>Stacklok, State of MCP in Software 2026, as reported by Digital Applied · all-software cohort</p><p>The same survey named security as the top barrier, ahead of implementation cost. That’s worth saying out loud in the room: MCP lowers integration cost, but giving AI access to business systems still needs approval gates, access controls and audit logs.</p><h3>The one-slide cheat sheet</h3><p>This is the table I now bring into these conversations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EoJ5KVwJGnTCuyqCJcIX4A.png" /><figcaption>Analogy comparison: REST vs. JSON-RPC vs. MCP</figcaption></figure><p>The last row is the one that defuses the original question. This isn’t “MCP <em>versus</em> REST”. It’s REST for your systems, JSON-RPC as the language, and MCP as the adapter that lets AI use both.</p><h3>What I do now</h3><p>That one question reshaped how I prepare for conversations with business leaders. Before any discussion about technology choices, I ask myself: what decision is this person trying to make, and what do they need to make it?</p><p>When a leader can reason about a technology on their own terms, the conversation shifts. They stop nodding along and start asking the questions that matter: “What does it cost to secure?” “Which systems would we connect first?” That’s when buy-in actually happens.</p><p>My rules for explaining technical choices to executives:</p><p>1. <strong>Lead with the why, not the how.</strong> Cost, speed, risk and leverage come before syntax.</p><p>2. <strong>Use one everyday analogy and stick with it.</strong> Switching metaphors mid-explanation loses people.</p><p>3. <strong>Show the math.</strong> Multiplicative versus additive cost lands instantly with finance minds.</p><p>4. <strong>Bring one or two real numbers.</strong> Adoption and backing from major vendors reduce perceived risk.</p><p>5. <strong>Name the catch.</strong> Being upfront about security builds more trust than a perfect pitch.</p><p>Next time someone asks you to explain a protocol, take them grocery shopping first. Save the curly braces for the engineers.</p><h3>Sources</h3><p>• <a href="https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation">Anthropic: Donating the Model Context Protocol and establishing the Agentic AI Foundation</a></p><p>• <a href="https://blog.modelcontextprotocol.io/posts/2025-12-09-mcp-joins-agentic-ai-foundation/">MCP blog: MCP joins the Agentic AI Foundation (Dec 9, 2025)</a></p><p>• <a href="https://www.digitalapplied.com/blog/mcp-adoption-statistics-2026-model-context-protocol">Digital Applied: MCP Adoption Statistics 2026 (Stacklok survey data)</a></p><p>• <a href="https://devops.com/2020-postman-state-of-the-api-reveals-api-investments-will-remain-strong/">DevOps.com: 2020 Postman State of the API</a></p><p>• <a href="https://www.infoworld.com/article/2338668/ai-is-catching-on-with-api-developers.html">InfoWorld: Postman 2023 State of the API findings</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eba9e4cd2d8b" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>