Sandipan Haldar

Applied AI Consultant
I write about AI Engineering stuffs here : @IITG

[Sandipan Haldar @sandipuns]

the only successful agents are Coding agents and deep research agents(somewhat ) .Agent UX pattern they have in common
1. Response streaming
2. Message Queuing
3. Interruption
4. Slash commands
5. Long running task with parrellel exection
6. Function call with Human approval

[Sandipan Haldar @sandipuns]

@theo @trq212 You can just use mdx for much fewer token burn

[Sandipan Haldar @sandipuns]

@raffichill Sick ui

[Sandipan Haldar @sandipuns]

@FarzaTV This is cool. I was thinking autogenerated callouts / tooltips for better guidance like a product onboarding tool

[Sandipan Haldar @sandipuns]

@karpathy @NirDiamantAI If you consider your coding agents as a compiler and your prompt as code then there needs to be some deterministic checks for the compiler to ground on. Your coding agent needs to be pure function.

[Sandipan Haldar @sandipuns]

@DevanshuXi @ExaAILabs @WilliamBryk really nice article !!

[Sandipan Haldar @sandipuns]

@Hxlfed14 This is very well written .

[Sandipan Haldar @sandipuns]

@andrew_r @zeeg people who hate MCP are usually taking about saving tokens/ context window - there is a Token-Interop tradeoff and that should be the conversation

[Sandipan Haldar @sandipuns]

@techgirl1908 This happens all the time. Inventing tests up so it can pass

[Sandipan Haldar @sandipuns]

A client asked me to deploy a RAG solution in their Azure cloud. Looking to use microsoft foundary tools. Even got myself a certificate for their course.
https://t.co/IqK5RIc973 #MSLearnBadge

[Sandipan Haldar @sandipuns]

@KillerShoaib__ Great job . Have you tried running some hard tests? What do you think are the limitations now ?

[David Soria Parra @dsp_]

Excalidraw in Claude. MCP Apps made by one of the main engineers behind MCP Apps: https://t.co/MxTFShG4Oe

[Sandipan Haldar @sandipuns]

@IsomorphicLabs You should fix your website

[Sandipan Haldar @sandipuns]

@_avichawla Why is it only for finance benchmark ? Would love to see it perform on some standard benchmark

[Sandipan Haldar @sandipuns]

We will be increasing seeing long running tasks get better this year.

[Cursor @cursor_ai]

Cursor now uses subagents to complete parts of a task in parallel.

Subagents lead to faster overall execution and better context usage. They also let agents work on longer-running tasks.

Also new: Cursor can generate images, ask clarifying questions, and more.

[Sandipan Haldar @sandipuns]

I’m working around detecting user struggle during onboarding workflows.
The idea is to capture fine-grained client-side interaction signals (form state changes, validation failures, retries, navigation context, timing) and then use an LLM to reason over those traces to infer where and why users are getting stuck.

I’m currently thinking about where this data should be captured and modeled:
raw DOM / browser-level events
or higher-level semantic events (e.g. “attempted submit → validation error”)

From your experience, what’s the right abstraction layer to collect this kind of interaction data so it's useful for downstream reasoning?

[Sandipan Haldar @sandipuns]

@zachbruggeman @tryramp @opencode @modal @CloudflareDev What's your cost for running this?

[Sandipan Haldar @sandipuns]

@brian_lovin Loved the website

[Sandipan Haldar @sandipuns]

@trq212 Can we use other LLM providers with claude agent SDK ?

[Sandipan Haldar @sandipuns]

@rork @RevenueCat would love to try

[Sandipan Haldar @sandipuns]

Think of it like the game Factorio*. You start by hand-crafting individual items (authorship), but to scale, you have to build an automated factory (orchestration) .You stop caring if you personally "made the gear" and focus on designing the system that ensures the gears are produced correctly at scale.

[Sandipan Haldar @sandipuns]

Even @karpathy says he’s never felt "this behind" as a programmer. If he feels that way, it’s a sign that the very definition of being technical has shifted under our feet . We’re moving from a world of writing code to a world of orchestrating intelligence

[Sandipan Haldar @sandipuns]

@karpathy The most important divide in 2026 isn't "engineer vs. non-engineer" it’s those who can delegate vs. those who can't
Whether you’re a lawyer or a coder, you’re now climbing the same skill tree: orchestrating uncertainty without losing authority