Sreenivas (@srmakam) header
Sreenivas (@srmakam) profile picture

Ex: KGeN (CTO), Google, Cisco | Engineering Leader | Author

Bengaluru, India
Joined March 2016
Photo by Sreenivas (@srmakam)Photo by Sreenivas (@srmakam)Photo by Sreenivas (@srmakam)Photo by Sreenivas (@srmakam)Photo by Sreenivas (@srmakam)Photo by Sreenivas (@srmakam)Photo by Sreenivas (@srmakam)Photo by Sreenivas (@srmakam)Photo by Sreenivas (@srmakam)Photo by Sreenivas (@srmakam)
The Udemy courses I valued most connected first principles with practical work. That shaped Inside AI Agent Systems: From Components to Platforms—a structured path through concepts, experiments and platform choices. udemy.com/course/inside-ai-a…
486
2/3 I use Codex Cloud and Gemini to review each other’s answers. It helps me challenge architecture decisions, catch missing details, and check the reasoning. Code review is another use case.
1
22
3/3 Frenemy also supports Claude and Kimi. One limitation in this setup is model selection. I can choose which agent to ask, but the receiving CLI selects the model.
12
1/6 How far have open-weight AI models come? Artificial Analysis’s Progress Index tracks the best open-weight model over time. The overall line rises substantially. What happens if we cap model size? artificialanalysis.ai/models…
1
46
5/6 The existing Score vs. Output Tokens per Task chart shows the trade-off. Each dot is a model at a given effort setting. At a similar score, the dot farther left got there with fewer tokens. artificialanalysis.ai/evalua…
1
12
1/3 A visual response from an agent can be easier to understand and act on than a block of text. As the question or task changes, the interface should adapt. I tried three ways to do that in my marketplace app, using the same evidence.
1
24
2/3 How much freedom should the agent have? • Controlled: select a prebuilt view • Declarative: compose agreed components • Open-ended: generate the UI code My takeaway: choose the freedom your task needs—and the behavior you can validate.
1
14
3/3 MCP Apps embeds UI in a compatible host. A2UI describes the view. AG-UI connects app and agent events. Different roles; they can work together. Walkthrough to help choose an approach: youtu.be/AFwt2MjEN6w Which approach fits your use case?
19
1/4 A support agent knows the marketplace. A delivery provider knows the shipment. How can support ask the provider to investigate without sharing its internal implementation? I explored that boundary with A2A in my marketplace agent application.
1
34
3/4 An Agent Card describes what the agent offers and where to call it. Discovery still needs a starting point: a known domain, registry or configuration. A2A connects to an agent; MCP connects to tools and resources. An agent can use both.
1
15
4/4 A2A helps independently managed agents work across team, platform or provider boundaries through a common interface. Not every agent-to-agent call needs it; agents within one app can use simpler approaches. Walkthrough: youtu.be/TNz8u-eqjGU
5
1/4 An AI agent has a valid credential. Can it read another buyer's order? That depends on the permissions enforced behind the tool. I tested MCP access in a marketplace agent to make this boundary concrete: whose authority is the call using?
3
1
82
3/4 The useful tests included denials, not just successful logins. My own buyer order succeeded; another buyer's order was blocked by the backend. The payout worker also faced eligibility and amount limits, and retries did not pay twice.
1
11
4/4 My takeaway: choose whose authority the tool call represents, then enforce ownership and business rules in the backend. A valid token alone is not permission for every action. Full video walkthrough: youtu.be/uxcTVYYaOxc
12