Mitul Sarvaiya

Building AI agents and Superintelligence.

[Mitul Sarvaiya @mitulsrv]

@TTrimoreau In between, I debug code and manually change if anything is needed.

[Mitul Sarvaiya @mitulsrv]

now I can share prototype directly and I believe it would close the gap between idea and shareable product.

[ClaudeDevs @ClaudeDevs]

Artifacts in Claude Code are now also available on Pro and Max plans.

Ask for an artifact, Claude writes the code, publishes it live to claude.a‍i, and updates it in real time while it keeps working. Pages are private to your account and fully self-contained.

[Mitul Sarvaiya @mitulsrv]

Let me go for walk now....

[Mitul Sarvaiya @mitulsrv]

@VaibhavSisinty have you tried it brother??

[Mitul Sarvaiya @mitulsrv]

@rdbuilds7 https://t.co/6UAM9hi4qW

[Mitul Sarvaiya @mitulsrv]

@billgnofficial https://t.co/6UAM9hi4qW

[Mitul Sarvaiya @mitulsrv]

@Gamingtronium I believe engineer B also generated the code but as he spent time on unit test, integration test, documentation and reviews so.

but think like, engineer A just vibe coded the feature but on the otherhand now engineer A becomes engineer B, and let the feature done.

[Mitul Sarvaiya @mitulsrv]

@claudeai I mean how!!!!!!!

[Mitul Sarvaiya @mitulsrv]

@claudeai
I just got weekly limit today suddenly in my account yet I didn't used that much tokens and in my current session, I have usage of 27%. Today, morning when I was working with Claude Code, I checked my weekly limit around 60%, so my question is within few hours, it is reached to 100%, is it possible?

I am on my MAX plan and so my team is also in MAX plan, everyone faced the same issue.

[Mitul Sarvaiya @mitulsrv]

@0xMovez looks so cool!

[Mitul Sarvaiya @mitulsrv]

@GoogleCloudTech done similar already, but thanks

[Mitul Sarvaiya @mitulsrv]

@narendramodi @BharatInnov2026 We are exporting these tech to abroad but what if we use the same in India and for India!

[Mitul Sarvaiya @mitulsrv]

@narendramodi I am in India and building AI development agency which build and trains LLMs, ML models and Deep Learning. I wish and I hope, we could help India and Govt to do AI Enablement. I tried a lot to contact to Govt. to do AI Enablement within it but never got an opportunity due to others.

[Mitul Sarvaiya @mitulsrv]

@socialwithaayan I think I need to read more about it.

[Mitul Sarvaiya @mitulsrv]

If you're hitting token limits, latency walls, or accuracy issues in multi-tool workflows, this pattern is worth looking at.

[Mitul Sarvaiya @mitulsrv]

An AI agent that calls tools one by one is not an agent.

It's an expensive middleman.
Here's what we changed, and what it actually looks like in production.

[Mitul Sarvaiya @mitulsrv]

The engineering moat is not in how many tools your agent can access.

It's in whether your agent is designed to think, or designed to execute.

The best systems know exactly when to do which.

[Mitul Sarvaiya @mitulsrv]

We read the AWS blog on PTC, tried it the same week.

We were already building agents with the Anthropic SDK on Bedrock, so we used the SDK-compatible proxy approach, same codebase, different execution layer.

No full rebuild. Just a smarter design.

https://t.co/9b1RBMv46p

[Mitul Sarvaiya @mitulsrv]

We maintain a Customer Account Health Review Agent for a customer on AWS Bedrock.

Before → 15–20 sequential tool calls, ~128k tokens, ~40s per review, NL aggregation errors.

After → parallel sandbox execution, ~13k tokens, ~8s per review, deterministic Python logic.

[Mitul Sarvaiya @mitulsrv]

Programmatic Tool Calling flips the design.

The model writes Python code that orchestrates all the tool calls inside a sandbox.

asyncio.gather() fires all sources in parallel. Python filters and aggregates. Only the final summary returns to the model.

Model is sampled exactly twice.

[Mitul Sarvaiya @mitulsrv]

In a standard agentic workflow, the model calls a tool → result floods the context window → model calls the next tool → repeat.

For a single account review: 15–20 tool calls. Every raw record passing through context. ~128,000 tokens. ~40 seconds.