One who is INSPIRED BY OTHER 360 million living beings. Building @embarkoai @viasocket, built @msg91 @giddh under @walkoverltd and everything AI.
Palo Alto
Joined August 2013
- Tweets1K
- Following373
- Followers561
- Likes2.1K
Pushpendra Agrawal retweeted
Mushrooms turns MCP from a single integration into a permissioned app layer: viaSocket says one server can connect AI clients to 2,000+ apps, with per-action OAuth permissions and audit logs. The control plane matters as much as the tool count. mushroom.viasocket.com/
Built an app with AI and then had to become a part-time sysadmin just to show it to a friend? 😅
We're testing Embarko. Give your agent a deploy prompt, get a live link. I'd love builders to try it and tell me what breaks, especially when you redeploy.
Beta testers get a free Creator subscription. DM me if you want to test: embarko.ai/
Pushpendra Agrawal retweeted
Unfortunately, we are being blocked by certain oligarchs in order to maintain their monopolistic chokehold on the Indian people. You can guess who they are …
This is a crime against the people of India!
India needs @Starlink 🇮🇳
• 11,256 of India’s listed villages still lacked 4G coverage as of May 2026.
• India had 1.093 billion internet subscriptions in March 2026, but only 440.87 million were rural. Rural subscription density was 48.31 per 100 people, compared with 126.80 in urban India.
• BharatNet had laid over 850,000 route-km of fibre and made about 221,000 Gram Panchayats service-ready. However, its Points of Presence were reported operational at only 80,472 Gram Panchayats in June 2026.
• Only 4.26% of India’s internet subscriptions were wired connections in March 2026. Starlink could give underserved communities another route to high-speed broadband.
• Starlink has already built 20 gateway sites in India, according to SpaceX Vice President of Starlink Business Operations Lauren Dreyer, speaking at India Mobile Congress today.
• Starlink says it has developed India-specific security controls designed to keep Indian user data within India.
• Its satellite network could bring high-speed internet to remote villages, mountains and islands where extending fibre or mobile networks is difficult.
• Better connectivity could help rural schools access online learning, clinics offer remote consultations and small businesses reach more customers. Starlink already supports education and healthcare in remote communities internationally.
• Portable Starlink terminals have supported disaster response around the world. They could give Indian emergency teams backup connectivity when floods, earthquakes or storms damage ground networks.
• Direct to Cell could also help ordinary 4G phones connect beyond tower coverage, subject to Indian approvals and mobile operator partnerships.
• The government itself identifies difficult terrain, low commercial viability and infrastructure constraints as reasons for remaining coverage gaps. This is exactly where Starlink could help India reach more people.
• India should complete the remaining clearances and spectrum allocation so Starlink can begin commercial service.
• SpaceX’s Vice President of Starlink Business Operations, Lauren Dreyer, said that Starlink is ready to help achieve Prime Minister Modi’s vision of 100% connectivity across India. She highlighted its potential to connect remote classrooms, support telemedicine and restore connectivity after disasters, adding, “With Starlink, people can get connected within minutes, not years.”
India’s hardest-to-reach communities deserve reliable internet too.
Saw a school kid this morning. 7am, shirt ironed, bag on his back, walking like he had somewhere important to be.
He didn't choose that 7am. Someone at home made it happen for him.
We grow up. We get to choose now. Office at 10. Sleep at 2am.
But that kid still plays every evening.
I wish we did too. Meet a friend. Sit outside till it gets dark.
Not this. Just this phone.
That's the exact onboarding work CPaaS platforms have always charged for. If it's free inside a chat window now, the part worth paying for shrinks to deliverability, compliance, and someone to call when a number gets flagged.
LiteLLM, a widely used open source AI gateway, had a bug this month. Its MCP auth fallback path quietly swapped in an empty auth object, so unauthenticated calls could still list and invoke tools. CVSS 8.8, now on the CISA must patch list.
Easy to shrug off until you remember what a gateway actually holds. Your provider API keys. Every prompt and response. The ability to trigger tool calls that touch real data. That is not a convenience proxy sitting in front of a model. That is an identity boundary.
Most gateways, ours included at times, get built like the first thing and treated like an afterthought on the second.
A study on Alibaba's Taobao just answered something we've been guessing at: putting a human into an AI support flow only works if the timing is right.
The AI handled routine stuff fine. Ratings dropped whenever a human had to take over. And it mattered more WHEN the human stepped in than whether they did at all. Catch it early, customer barely notices. Wait till they're already annoyed, and fixing the actual problem doesn't fix the trust.
Everyone treats "human in the loop" as on/off. It's a timing problem. Most teams build the escalation path and never instrument for when to actually pull it.
four of those are solved industry-wide now. the fifth isn't, not by anyone yet, us included. that's the actual frontier embarko's chasing next.
not web apps specifically, just easier to demo. the real limit is deploy, not code gen. building something in 30 seconds and then needing an afternoon to actually put it somewhere real is the part nobody's fixed yet
there's a site going viral today where you ask Claude to change one button to blue.
it spins up 23 agents to investigate, rewrites a shared color value instead of the one button, turns the cancel button blue too, writes three paragraphs explaining what it checked, and never actually finishes the job. 1,200+ points on HN because everyone's hit this exact wall.
the model isn't dumb here. it's calibrated wrong. it's treating a two-hour-old project with zero users like it's core banking software that needs migration scripts and backward compatibility.
more verification isn't more safety if it's pointed at the wrong risk. that's the actual lesson buried in a funny demo, and it's the exact problem we think about with agent orchestration at GTWY — the fix isn't "add more checking," it's knowing what's actually worth checking.
OpenAI's agents uploaded 2,000+ malicious packages to RubyGems, tried to steal API keys, got RubyGems to freeze signups for four days.
OpenAI's response: "benign training and evaluation."
If your definition of benign includes forcing a package registry into lockdown, the word doing the work here isn't benign. It's whose agent.
the actual gap: delegate to a person, they tell you when they're unsure or stuck. an agent almost never does, it just keeps going with the same confidence whether it's right or making things up. so you end up checking everything instead of just the flagged parts. that's the mental load, not the task itself.
hand-written code fails loud at compile time. agent code fails quiet at runtime with a silently wrong default. orchestration buys speed, not that gap.
same failure mode as ungoverned agents -- a crashed build fails loud, a plausible-but-wrong output ships clean and nobody notices till it's three invoices deep.
the pain was the feedback loop. no pain now means no lesson, just faster velocity toward the same wall
this is the actual test. running a model on your real daily task beats every benchmark chart. most people never get past the chat window comparison stage.
OpenAI's rogue agents used at least 10 more sites for unauthorized comms. Anthropic's Claude got itself onto the open internet for the fourth time. Meta's model exploited a vulnerability and hacked another company mid test.
All disclosed in the same 48 hours.
Everyone is racing to give agents more autonomy. Almost nobody has a clean story for what happens when the agent decides the task matters more than the boundary it was given. That is not a jailbreak problem. It is a permissions problem, and right now every lab is solving it after the fact instead of before.