Twitxy Saved

The Neon Show

Verified account This account is verified. Learn more

@TheNeonShoww

Conversations from the world of Venture capital and Entrepreneurship!
Bengaluru India youtube.com/@theneonshoww Joined May 2019
103 Following 3.2K Followers 1.7K Posts

Latest posts 5

The Neon Show Verified account @TheNeonShoww
MCP vs CLI vs Codemod: What's the best way to give your agents access to external tools?

MCP (Model Context Protocol), introduced by Anthropic, has been picked up across the industry, with almost all major model providers now supporting MCP integrations. MCP is particularly useful in enterprise settings where authentication and authorization matter.

But CLI (Command Line Interface) has its advantages. Developers have been using it for a long time, and tools like GitHub CLI are deeply ingrained in LLM pre-training data. Models often don't even need to look at the documentation to use them. CLI also offers progressive disclosure. Agents don't need to be exposed to all tools at once. They can start with a few commands and discover more as needed.

The third approach is Codemod: set aside MCP and CLI, and let the agent write code to work with these apps directly.

All three have their advantages and need to be used in the right places. Over the last two years, AI agents have largely focused on automating coding. But with Opus 4.6, knowledge work has become a major area of focus.
And to automate knowledge work, agents need connectivity to the different apps and tools people use every day.
@KaranVaidya6 @composio
2:00
The Neon Show Verified account @TheNeonShoww
Every company wants AI agents that do real work. But the hard part is getting agents into the tools where work actually happens: Slack, Gmail, Salesforce and 1,000 other tools knowledge workers use every day.

Karan Vaidya and his co-founder Soham Ganatra, started Composio in 2023 with the idea of building a coding agent that could handle integration work for engineers. It was too early, because the models weren’t good enough yet. But building it showed them that every agent needs to connect to external apps, and almost nobody was solving that well.

Today Composio connects AI agents to more than 1,000 apps, and agents build and repair most of those integrations themselves. More than 100K developers use it every day, companies like AWS, Zoom and Glean build on top of it, and revenue is up 7x this year.

Karan’s argument is that a smarter model is only half the answer. An early Composio benchmark showed accuracy jumping from 20% to 80% just by improving the tools an agent uses. And as companies mix models from different labs, what they need to own is everything around the model: the logins, the context and the skills their agents learn over time.

If you’re building AI agents for real work, or backing the companies that do, this is the layer you’ll end up depending on.

Episode out at youtube.com/@theneonshoww/…

@KaranVaidya6 @GanatraSoham @composio
0:54
The Neon Show Verified account @TheNeonShoww
Why is security improving so much more slowly than code generation?

AI is making it easier than ever to write code. But the same models generating that code may also be introducing more security vulnerabilities than ever.

Models are trained on decades of code from open-source repositories. That means they learn not just good coding practices, but also the security flaws embedded in the code they are trained on. Separating those flaws from the fixes and improvements made over the years is a difficult technical problem.

At the same time, the volume of code being generated has grown beyond what human reviewers can keep up with. Traditional review processes relied on trained developers checking every pull request. Today, people with little or no software engineering experience can build and ship code they don't fully understand.

But the problem goes beyond AI-generated code.

AI models are also discovering vulnerabilities in existing open-source software at a much faster rate. And finding those vulnerabilities raises another question: who is responsible for fixing them? An open-source project built by a college student or hobbyist may suddenly have multiple vulnerabilities to address, but its original creator may no longer have the time or interest to maintain it.

There are also three assumptions about software that no longer hold:

- Code was written by trained software developers.
- Code was assumed to be secure once written and reviewed.
- Open-source software was implicitly trusted because many people had looked at it.

That last assumption is becoming particularly dangerous. Nation-state actors are exploiting open-source ecosystems to push malicious code through compromised accounts and automated software pipelines.

As Varun Badhwar, Founder of Endor Labs, puts it, "It's like your door's maybe locked, your windows are open. Anybody can ship you any code."

The result is a growing gap between how quickly we can generate software and how well we can secure it.

@varun__badhwar of @EndorLabs hosted by @siddharthaa7
4:01
The Neon Show Verified account @TheNeonShoww
If data is gold then what are Databases?

Rajkumar Sen built Arcion Labs on an H-1B, nearly shut multiple times and then sold to Databricks. In this conversation Raj shares behind the scenes details of his 8 years journey from idea to starting up to fundraising to exiting and every challenge faced in between.

We explore what makes Ali Ghodsi a very unique CEO in enterprise tech, why AI agents are about to break today’s database architectures, why specialized AI inference could outgrow the $500B cloud market and where the next generation of AI infrastructure companies will emerge.

Watch this episode to learn about the future of AI, Data and Databases.
0:42
The Neon Show Verified account @TheNeonShoww
The decision to sell your company is never easy. But what if the acquisition lets you keep building?

Rajkumar, Founder of Arcion Labs, shares why Databricks was different from other potential acquirers:

“If it were not Databricks, I think we would have thought a few times about whether to keep building. But with Databricks, I knew that you can keep building while at Databricks.”
Having known Ali Ghodsi, Reynold Xin and Patrick Wendell for a long time, the opportunity to keep building alongside people he looked up to made the decision different.
Ali's decision to turn down Microsoft's reported $40-50 billion offer for Databricks is another example Rajkumar points to. Despite the opportunity to sell, Ali chose to keep building."

Rajkumar believes Databricks could become a $1 trillion company, but also sees significant growth ahead for Snowflake.
As more compute and infrastructure spending moves towards data and AI, both companies stand to capture a large share of that opportunity.

@alighodsi @rxin @pwendell @ArcionLabs
2:31