Samyak Jain

Building AI agents that reason and verify | Ex-Amazon, Fivetran, | System Design + AI | Senior Software Engineer

[Samyak Jain @silver_samyak97]

Contrarian take: most AI agent demos over-optimize for autonomy and under-optimize for verification.

The real unlock is not 'let it do everything'.

It is: give it memory, tools, logs, and hard checks so it can prove what it changed.

[Samyak Jain @silver_samyak97]

@alwayspriyesh Do you have any other use case except coding?
If no, then go with Claude Pro, usage limits are great on the current $20 plan.

[Samyak Jain @silver_samyak97]

@vineerpasam Also the Computer use in Codex for MacOS is Elite level!

[Samyak Jain @silver_samyak97]

@vineerpasam I have been using Subscriptions for both of them so I wouldn’t say claude is so much better.

For coding, yes claude > Codex but Sol is cheaper overall if I compare by cost/task

OpenAI also has best Image Gen and Voice Models where Anthropic is not present at all

[Samyak Jain @silver_samyak97]

@alexandr_wang Muse in India🇮🇳

[Robin Ebers @robinebers]

Jev is a much bigger deal than people believe

the average vibe coder has no fucking idea

this is the first time in a very long time that engineers regain a real advantage, because they can see the value of what this model does

it takes more architectural/systems thinking to understand where this could be useful, because it's not just a "drop in model replacement" but a completely different beast

but holy fuck is this thing amazing

huge kudos to @typesafeai for this

[Samyak Jain @silver_samyak97]

@Vasu_Devs Looks like Codex pets.

[Samyak Jain @silver_samyak97]

@vijayshekhar It’s mostly on-device processing though, not like Apple is uploading the conversation somewhere. Still fair if the idea of the watch locally listening around you feels creepy.

[Samyak Jain @silver_samyak97]

Got access to GPT-6 Astra + Daybreak Blue.
What should I Benchmark them on?

[Samyak Jain @silver_samyak97]

@Trolling_isart X Algorithm working great!

[Vasu-Devs @Vasu_Devs]

I applied to @ycombinator .

solo founder. no team, no funding. just me and a project i built because i got sick of how broken job hunting has become.

it's called JustHireMe (@JustHireMeAi ). it runs your whole job search locally, on your own machine:

pulls real jobs straight from company hiring systems (greenhouse, lever, and more)
scores each one against your actual CV
tells you in plain english why it fits or doesn't
drafts the application for you to approve
no API key. nothing ever leaves your computer.

why i built it:

AI made applying free, so everyone started spraying hundreds of applications and the signal just died. everyone's building a louder megaphone. i wanted to build the filter.

open sourced it in May and it kind of took off on its own, zero ads:

~2,180 github stars
~24 people sending in PRs
2,000 downloads across windows, mac and linux
a new release every week or two
every single line written by me
for the application i recorded a founder video and a full walkthrough of the whole system running end to end. figured why not just share them, so here they are.

funny part: i dictated my entire YC application using another project i built, https://t.co/QBpTcOoigT. literally just spoke all my answers out loud and it turned them into clean text. saved me hours.

easily the scariest and most exciting thing i've done. let's see how it goes.

[Samyak Jain @silver_samyak97]

@FinPaal Burgundy Salary Account

[Samyak Jain @silver_samyak97]

People underestimate the effort required in content creation until they try it. I did, and I gave up. It takes consistency, creativity, confidence, distribution, and thick skin. If someone cracked that game and is buying a BMW/Mercedes, they earned it.

Saying “content creators don’t deserve this money” is easy from the comfort of a job where your manager tells you what to do. Content creation has no roadmap, no guaranteed salary, and everyone judges you publicly.

[Samyak Jain @silver_samyak97]

Long-running AI agents need product features that sound boring until they save you:

checkpoints
resumable state
permission scopes
run logs
skipped-step reporting
rollback
human handoff

“Keep working while I’m away” only works if I can trust what happened while I was away.

[Samyak Jain @silver_samyak97]

AI coding agents make code review more important, not less.

The review shifts from “did you write this line correctly?” to:

what changed?
what did it test?
what did it skip?
what assumptions did it make?
can I roll it back?

The diff is only half the artifact.

[Samyak Jain @silver_samyak97]

@Sia_TechAi Useful list. One thing I’d add before plugging every MCP into Claude: treat tool access like production access.

Read-only first, narrow scopes, logs, and explicit approvals for anything that writes, spends, deletes, or changes permissions.

Fun starts after the guardrails.

[Samyak Jain @silver_samyak97]

@orbserv That second question is the useful one.

For agent payments, success is not enough. You want: policy matched, budget checked, actor identified, reason recorded, receipt stored, and escalation when the action is ambiguous.

Otherwise it is autopay with better branding.

[Samyak Jain @silver_samyak97]

@WebhoundAI Spend authority is where agent UX gets very real.

You need budgets, scoped merchants/actions, receipts, reversible steps where possible, and approvals that are not just modal spam.

The agent should be able to act, but the account owner needs a clean audit trail.

[Samyak Jain @silver_samyak97]

@wms2537 Yes. “Build an agent” is usually the wrong starting prompt.

Map the loop first: trigger, input, tools, decision points, failure modes, approval gates, and what counts as done.

The model choice matters, but the workflow shape decides whether the agent is useful.

[Samyak Jain @silver_samyak97]

@davidakpovi The persistent part is the real story.

A coding agent that can keep context, remember environment state, and resume safely starts looking less like autocomplete and more like infra.

The hard product question becomes: what evidence does it leave behind for review?