Vivek (@vivek96_) header
Vivek (@vivek96_) profile picture
New Delhi, IN
Joined March 2014
Photo by Vivek (@vivek96_)Photo by Vivek (@vivek96_)Photo by Vivek (@vivek96_)Photo by Vivek (@vivek96_)Photo by Vivek (@vivek96_)Photo by Vivek (@vivek96_)Photo by Vivek (@vivek96_)Photo by Vivek (@vivek96_)Photo by Vivek (@vivek96_)Photo by Vivek (@vivek96_)
i don’t even remember when i stopped using Notion 🥀
35
Winter is coming 🥶
1
9
153
every time it dips 15%, i'm buying.
41
our CTO with his $200/mo intern at 4am
2
7
102
just 4 folks moving fasttt. apparently i’m the one who needs to catch up 😭 fix the agentic revenue leakage today: agentrail.co
1
4
175
People say I care about littering so much, idk why they keep asking me that.
1
7
121
Vivek retweeted
"How old are you?" me:
318
2,911
203
31,165
750,296
"it's done" is at least something you can check. the runs that worry me are the ones that just stop. no done, no error, nothing to review
If you ship with AI agents, "it's done" is the most expensive sentence you'll read. Here's how I check agent work after 6 apps shipped without opening Xcode: 1. Let the model review its own code, then have a different model review it again. A model checking only itself just nods along. 2. Judge against your own written rules. Never ask "is this good?" 3. When the two models disagree, bring in a third. The disagreement is where the bugs hide. 4. Measure cost in follow-up prompts, not price per run. A cheap run that needs 8 fixes isn't cheap. 5. "Done" means it works and a real person can use it. Check that part yourself. Your job isn't writing code anymore. It's context, constraints and checks, and knowing when the checks aren't enough. If you want a head start, try @garrytan's gstack, especially the /plan-eng-review skill. It reviews the plan before any code gets written.
2
1
62
who came back and who didn't is the page i'd open first. runs and tokens tell you the agent worked, not whether anyone wanted it again
Your agents have users. Now you can actually see them. Who’s active this week, what’s running right now, who came back and who didn’t, and a page for every person.
2
89
everyone's asking about sales. i'm stuck on step one: the agent opened a google ads account by itself. signup is where most agents give up in our data
I asked Replit to grow my app It spun up a Google Ads account on its own Started running campaigns And has gotten 10,000s of impressions for me with paid ads in the last 48 hours. AGI for making $$ is here.
41
700 sites as apis means agents stop clicking through pages built for people. the business portals are the interesting part. that's where the login wall is
Replying to @pomeradoai
@pomeradoai now supports APIs for over 700 sites - ecom (amazon/target/etc) - food delivery - flights and taxis - business portals people were making so many APIs we had to redesign the dashboard:
2
52
November 20. Stay tuned 👀
1
2
6
84
an agent-only brand is a fun bet. the catalog gets it found, checkout is where it's won. in the 4M+ agent events we looked at, the #1 failure was the login, not discovery
with agentic commerce, there will be a $100m viral brand soon that you can only buy through an agent. It will have no website. Only a shoppable, enriched catalog connected to a Muse, Instinct, Grokbot, etc. You can only buy it through talking to an agent. Will be an operator like Hudson from Comfrt, they’ll be the Hudson of playing the agent game. If this is something you are trying to do or build, DM me. We have the infrastructure to get any brand’s catalog 100% agent-ready and most favored by agents.
25
561 downloads in week one for an mcp that lets any agent reach private models. private mode as the next update makes sense, coding agents send a lot of code through these
561 downloads. First week on npm. 🫡 We shipped offrouter/mcp so any agent could reach private models. You showed up. So today it got its biggest update yet: private mode, where your coding agent’s messages are sealed before they leave your machine. Week two starts now 👇
24
execution getting cheap moves the bottleneck to whatever still needs a human in the middle. a lot of that is just software that only lets a person log in
The cost of execution in software is approaching zero. In that world, super helpful to have the agents steered by multiple humans able to coordinate directly without a human in the loop. Lots of problems from the old world just aren’t a thing anymore (like tickets)
27
45% of last week's axios downloads still had the bug, 2 months after the fix shipped. "we shipped a fix" and "people are running the fix" are two different numbers. good to finally see the second one
New on @cvedotnow: affected downloads 📉 A fix only helps if people actually install it. CVE pages for npm + Rust packages now show what share of last week's downloads were affected versions. Example: 45% of axios downloads last week were a version w/ a known prototype pollution bug. That's 77M downloads, 2+ months after the fix shipped 😬 cve.now/CVE-2026-101908
1
24
42 of 100 real sites worked on warm replay. so most of the web still has no clean way in for an agent. the sites that ship a documented api or an mcp don't need to be reverse engineered at all.
Most AI agents use websites the slow way. They open a browser, load the full page, read thousands of tokens of HTML and click around. Every agent repeats the same work on the same site. There is a faster way: call the site's internal API (the hidden endpoints the page itself uses to get its data). You get clean JSON back, with no page to read. But there is a catch. These APIs are not documented. Someone has to find them first, work out the inputs, and handle tokens and logins. And they can change without warning. A paper from the Unbrowse team ("Internal APIs Are All You Need") solves this with a shared registry of learned routes. The idea is simple: learn a site once, then replay it. The agent asks for a task, like "top stories on Hacker News". Unbrowse searches your private tools, then a public registry of sites already learned. If there is a match, it replays the site's own HTTP requests. No browser. If not, the agent does the task once in a recorded cloud browser, and the traffic is compiled into a one-call tool. The system works out by itself which request fields are inputs and which are fixed. How it works: → It records two sessions with different inputs and compares them. The fields that change are the inputs. → It traces tokens (like CSRF values) back to the response that produced them, so a multi-step flow replays as one call. → A run counts as success only when the result is verified. HTTP 200 alone is not enough. → Only read-only routes with no secrets go to the public registry. Logins and writes are never shared. Results in the paper across 94 production domains (self-reported, compared with Playwright): • 950 ms mean vs 3,404 ms with browser automation • 3.6x faster on average, 5.4x at the median • 90–96% lower cost per task • ~8,000 tokens of DOM down to ~200 tokens of JSON The team's newer test on 100 real sites is more modest: 42 worked on warm replay, 2 gave wrong results, and the rest failed or could not be learned. What I find interesting as someone building agents: the browser becomes the fallback, not the default. But only the CLI and SDK are open source. The service that does the work is hosted, so check that before you depend on it. Links in the reply 👇
3
54
years of getting google to notice us, and the agent stops at a form built for a person. it's the #1 failure in the 4M+ agent events we looked at. the login, not the model.
spent years getting google to notice us, only for an AI agent to get stuck on our signup page 🫠
3
59
the planes that never came back are the agents that never picked your tool. no log, no error, the goal just went somewhere else. good one prathmesh
day 8/ of mcp learnings survivorship bias with mcp, it’s easy to treat tool call volume as a proxy for users succeeding- and to prioritize the most-called tools (i made this mistake last year at @asana) in ww2, the us military wanted to add armor to its bombers- so they studied the planes that came back. they saw bullet holes clustered on the wings, and planned to armor those spots they were only looking at survivors though! planes hit in the engines didn't make it back tool call volume can skew a few ways: 1/ model never selects your tool! you won't have a log for this- user goal is underrepresented 2/ model calls your tool multiple times without delivering user value- every call is a server log w/ a success status code- user goal is overrepresented in scale’s mcp-atlas benchmark, 40% of o3 pro (old, I know)- failed tasks had no tool call at all don’t use raw tool call counts as a proxy for successful mcp server usage! pointing you back at my prev tweet on the "user value chain", request flow of: connection → discovery → tool selection → tool call → tool response → user value tool call volume only counts the middle more on this in my effective mcp blog: mcpjam.com/blog/effective-mc…
1
32
the login is where most agents die, not checkout. so identity before payments is the right order. shopify turning it on for every merchant is the big part. the long tail of saas with a human-only signup is the real test
Websites have spent twenty years building walls against bots: CAPTCHAs, rate limits, user agent sniffing, the works. Today Sierra tore a door into that wall and called it Poppy. Personal Agent Protocol 1.0 launched this morning as the identity and permission layer for agents that shop, book and pay on your behalf. It doesn't replace the standards already in the room, it wires them together: OAuth decides what your agent can and can't do and lets you change that access later, MCP and OpenAPI handle the actual calls, and a commerce extension co-authored with Shopify hands off to UCP, Google's checkout protocol, once an agent is cleared to shop. Poppy answers one narrow question before any of that happens: who is this agent, and what did the human actually authorize. The adoption numbers are the part worth sitting with. Shopify turned Poppy on for every merchant on its platform in one move. Meta's Muse, already set up by more than 10 million people, now checks for a Poppy file on every site it visits by default. More than 100 companies have signed on since the original eight design partners, Delta, Salesforce, Twilio, Wayfair and Zoom among the latest. Mercury is routing agent initiated payments through existing account approval rules. Turkish Airlines lets an agent search fares and flag an upgrade before the human gives final approval. None of that is a demo. It's live. Step back and this is the same fight Visa's Trusted Agent Protocol, Mastercard's Agent Pay, Stripe's Agentic Commerce Protocol and Coinbase's x402 are already having, just one layer up the stack. Those four are arguing over how an agent pays. Poppy is arguing over whether an agent gets to show up at all, and under whose rules. My read: the payment rails will matter less than everyone building them thinks. Whoever owns identity and consent owns the actual chokepoint, the same way OAuth quietly became the thing every app had to go through once third party logins got common, not the payment processor sitting behind it. What's still missing is anyone outside Sierra's own design partners proving this holds up against an agent that lies about what it's allowed to do. A hundred partners signing a spec is not the same as a hundred partners surviving the first real incident. teapawt.org/SierraPlatform/status/…
1
27