Pratyus Patnaik

An immigrant’s laptop. Only in America 🇺🇸

It’s official…welcome to Snowflake, @natomalabs! We’re bringing Natoma’s enterprise MCP platform for AI agents to Snowflake, making it easier to securely connect AI to tools like Slack, Jira, email, calendars and workflows through Snowflake Intelligence and Cortex Code. More info ICYMI: bit.ly/4frxZiq

We’re excited to share that @Snowflake has signed a definitive agreement to acquire Natoma! Natoma was founded on a simple belief: AI agents will transform how work gets done inside enterprises, but only if organizations can securely connect, govern, and control how those agents access data, use tools, and take action. Together with Snowflake, we’ll help enterprises bring together data, models, and control — enabling agentic AI that is trusted, connected, governed, and built for action. Read more from our founders: natoma.ai/blog/natoma-sn… Snowflake blog: snowflake.com/en/blog/snowfl…

Invest and sign a definitive agreement in the same quarter? Well, this is a new one! Congrats to @pratyus and the @natomalabs team on entering into a definitive agreement with @Snowflake. The demand for agentic AI infrastructure is moving fast. natoma.ai/blog/natoma-sn…

Today we announced our intent to acquire Natoma, the enterprise MCP platform. Once closed, @Snowflake users will be able to enrich their Snowflake data with critical application context and take action directly from Snowflake Intelligence and Cortex Code. One platform for every data and workflow need. Send emails, summarize Slack threads, check calendars, and open Jira tickets, all without leaving Snowflake. Snowflake already governs your data. With Natoma, we'll govern AI actions too.

The release candidate for MCP 2026-07-28 is out. The protocol is now stateless: no handshake, no session id, any request can hit any server instance. Plus extensions as first-class (MCP Apps, Tasks), auth hardening, and a proper deprecation policy so we don't have to do this again. blog.modelcontextprotocol.io/posts/2026-07-…

200K MCP servers on public IPs. STDIO transport runs any OS command it gets. Anthropic says it's by design. 9 of 11 registries accepted a malicious test package without review. This is not a security failure. It is an inventory failure. venturebeat.com/security/mcp-s…

Three paths to connect agents to external systems: - Direct API: agent writes HTTP in a sandbox. Fine for 1 agent <> 1 service. Breaks at scale. - CLI: agent runs shell commands. Works in local envs with a filesystem. Can't reach mobile, web, or cloud. - MCP: agent connects to a server over a standard protocol. Built for cloud-hosted production agents. Mature integrations ship all three. MCP is the one that compounds. One remote server reaches every compatible client (Claude, ChatGPT, Cursor, VS Code) across every deployment environment, and gets more capable as new protocol extensions land without you shipping anything new.

Your AI agent needs real tools to be useful. Giving it access to everything is a security problem. NVIDIA NemoClaw isolates compute. Natoma governs tool calls. Zero service credentials inside the sandbox.

@Jason I’m in

We started an AI founder twitter group... reply with "I'm in" if you're a founder and want to be added

Past 50 MCP tools, agent accuracy drops. By 200+, you lose 1 in 7 queries and token costs spike 10x. Our engineer @KarnikShreyas built the fix: search-then-execute with hybrid retrieval. 97.5% recall at enterprise scale. Constant context cost. Full write-up: natoma.ai/blog/the-invis…

🚨 CRITICAL: Active supply chain attack on axios -- one of npm's most depended-on packages. The latest axios@1.14.1 now pulls in plain-crypto-js@4.2.1, a package that did not exist before today. This is a live compromise. This is textbook supply chain installer malware. axios has 100M+ weekly downloads. Every npm install pulling the latest version is potentially compromised right now. Socket AI analysis confirms this is malware. plain-crypto-js is an obfuscated dropper/loader that: • Deobfuscates embedded payloads and operational strings at runtime • Dynamically loads fs, os, and execSync to evade static analysis • Executes decoded shell commands • Stages and copies payload files into OS temp and Windows ProgramData directories • Deletes and renames artifacts post-execution to destroy forensic evidence If you use axios, pin your version immediately and audit your lockfiles. Do not upgrade.

3 bits. Zero loss. ~6× compression. Outperforms vector search at Google scale. Bigger model isn't the frontier anymore - better & efficient is.

Thank you, @united!

We just added Skills to Playground by Natoma. Skills are everywhere - buried in repos, threads, and docs. No structure, no easy way to know if something's worth using before you try it. We fixed that. 🧵

Your developers installed an AI package your IT team has never heard of. Today it was found stealing credentials.

While everyone's busy at #RSAC. This just dropped: github.com/BerriAI/litell… LiteLLM (v1.82.7 & 1.82.8) is compromised via a malicious .pth file - runs on Python startup, no import needed. Steals creds. Exfiltrates silently. If you use LiteLLM, DSPy, or GraphRAG - check NOW! (CrewAI / Google ADK → verify integrations) Stay safe out there!

AI agents are moving from experiments to real enterprise workflows. But one of the biggest challenges isn't the models - it's connecting agents to the systems your business runs on. We've been working with @1Password on their Unified Access launch to make this simpler. 🧵

We automated testing. We automated builds. We automated deploys. But demos? Still screen recordings and prayer to the demo Gods! So I built Argo. Script your demo → generate the video. 🎬 npx argo pipeline my-demo github.com/shreyaskarnik/…

The browser defined the human era. A new orchestration platform will define the era of intelligence!

An agent isn't a prompt. It's a distributed system with a probabilistic core. Whether operating inside the enterprise or acting on our behalf in the real world, agents will require identity, access, dynamic policies, skill boundaries, context governance, observability, and kill switches

Browsers were built for humans. You come to a page, log in, click around, and do your job. Soon, agents will be doing the work. And agents don't need UI or dashboards. The next browser won't render HTML - it will pipe data, context, and actions. The next "user-agent (UA)" will look more like a hypervisor or Kubernetes: an orchestration plane for AI systems.