Amit

Energy sector. Building AI agents with memory and audit trails.

[Amit @iamitp]

@hthieblot Google Reader

[Amit @iamitp]

The AI adoption question I keep coming back to: can this workflow survive a Monday morning in a real institution? If not, it is still a demo.

[Amit @iamitp]

Hey #Codexers what do you think?

[Amit @iamitp]

Still not resolved it's been over a week! Unsubscribing.

[Amit @iamitp]

So @WisprFlow is still not working on Mac @tankots #wisprdown

[Amit @iamitp]

So @WisprFlow is still not working on Mac @tankots #wisprdown

[Amit @iamitp]

Saved hours compound quietly. Announced budgets do not. I track what actually changed in my week, not what got approved, and the gap between the two numbers is always instructive.

[Amit @iamitp]

So @WisprFlow is still not working on Mac @tankots #wisprdown

[Amit @iamitp]

@tankots

[Amit @iamitp]

I, alternatively, just made them into MCPs

[Nityesh @nityeshaga]

I just turned one of our most token hungry skills into a dynamic workflow - now it uses upto 5x fewer output tokens while spawning more subagents!

This means significantly lower AI costs for this job while increasing the reliability of the workflow purely from changes in the harness level. If this was running this on API pricing, savings would be massive - while increasing the reliability!

Anthropic's insight of replacing an orchestrator from a non-deterministic AI agent to deterministic code written by an agent is such a simple and brilliant insight, it's blowing my mind! 🤯

So if you have tried to create skills to let your AI agent work do complex work, especially non-technical, it's time to ask your Claude to turn it into a Dynamic Workflow and see if that improves the outcomes.

[Amit @iamitp]

So I am trying to solve it by build a "me-emulator" with a local model

[Nathan Baschez @nbaschez]

This is a bigger idea than animations

A real blind spot for AI right now is it’s much better at doing what you say you want than it is at telling you what you should consider

Eg let’s say you have a realtime data processing task that has to handle heavy load. If you just describe the task the LLM will make a little web app to do the job. But an experienced engineer would ask questions and might suggest something like Kafka depending on your answers.

Feels at least partially solvable for the labs, but currently there is a lot of alpha to having an extensive vocabulary of patterns and knowing when and how to use them.

[Amit @iamitp]

@ameyakarve Nice nice!

[Amit @iamitp]

@ameyakarve Wired in transaction-level accruals?

[Amit @iamitp]

Generic assistants plateau because they optimise for the average user, and you are not average. The gap is not closed by a bigger base model. It is closed by your corrections. The base model is rented. The captured judgement is the asset you own.

[Amit @iamitp]

I automated my finances to save time. The real return was different: it caught the quiet leaks a human skips. Crept up subscriptions, duplicate charges, a standing instruction paying for a cancelled service. Individually trivial. Together, a steady drain.

[Amit @iamitp]

@thsottiaux CLI -> Remote control

[Amit @iamitp]

The conference version of Indian AI adoption is a chatbot and a panel. The real version is in reconciliation queues and messy document pipelines, built frugally by teams who measure success in hours saved. Look at the back office, not the keynote.

[Amit @iamitp]

Running ~50 automated jobs solo on one Mac Mini taught me the real lesson of personal AI infrastructure: capability got cheap, curation got expensive. The binding constraint is no longer producing output. It is reliably deciding what to keep.

[Amit @iamitp]

Ok been using Opus 4.8 and Ultracode for past 10 hours and I miss my GPT 5.5. Waiting for weekly reset.

[Amit @iamitp]

@Dimillian Will it pickup CLI Codex sessions now? For remote control

[Amit @iamitp]

@thsottiaux Bro please add remote control for CLI already