Ajeet

Software engineer. Sharing what I learn about building a tech career, one lesson at a time. Books I read, mistakes I made.

[Ajeet @AjeetThinks]

Sometimes your AI doesn’t need to write a novel. It just needs to answer: retry, escalate, or stop?

OpenAI’s Decisions API returns structured choices, scores, and probabilities, targeting decisions up to 10× faster than the Responses API.

[Ajeet @AjeetThinks]

Imagine assigning 1,946 AI agents one gene each, letting them read the research then combining their findings in 32 minutes.

OpenAI, PerturbAI and Scripps Research did exactly that.

Multi-agent systems are starting to look less like chatbots and more like distributed computing.

[Ajeet @AjeetThinks]

AI agents can generate code faster than ever. They can also generate infrastructure load faster than ever.

GitHub moved 99.99999% of its repositories to Azure as Git traffic surged.

The bottleneck moves. Good system design anticipates where it goes next.

[Ajeet @AjeetThinks]

GitLab CVE-2026-110264 has a CVSS score of 10.0.

An integer overflow in a regex compiler could allow unauthenticated remote code execution.

We patch application dependencies religiously, then forget the platform deploying our code.

That’s a dangerous blind spot.

[Ajeet @AjeetThinks]

Giving an AI agent root access because it promised to be careful is peak software engineering comedy.

NVIDIA OpenShell puts the guardrails outside the agent, with sandboxing and policy enforcement.

Your agent can be smart. Your permissions should be smarter.

[Ajeet @AjeetThinks]

Imagine spending weeks hardening your API auth, then a PHP runtime bug lets an unauthenticated request execute code before your app even starts.

[Ajeet @AjeetThinks]

Github is down 😩

[Ajeet @AjeetThinks]

@KomalWCode I'd inspect planning errors, lost state, and missing postcondition checks.

[Ajeet @AjeetThinks]

@anah_sahh The next big AI improvement might not be smarter answers, but knowing when to stop talking.

[Ajeet @AjeetThinks]

@LeoBuilds_ My model-switching finger is waiting for a better reason than a new version number.

[Ajeet @AjeetThinks]

@AnkanXplorer 6 digits gives 1 million combinations. But the real security depends on expiry and rate limits, not just the number of digits.

[Ajeet @AjeetThinks]

@codvynn The new utility class that solves a problem nobody had.

[Ajeet @AjeetThinks]

feature works on day 2 and sits in staging until the review on day 10. the scrum guide never asked anyone to wait. most teams hold releases out of habit, then blame the framework for being slow.

[Ajeet @AjeetThinks]

reading other people's code used to be a punishment. now it's the job.

[Ajeet @AjeetThinks]

letting the same model review the design it just wrote is approving your own PR. a june arXiv setup has gemini draft and gpt-oss review across 12 design sections. i'd steal that before trusting one model's "looks good".

[Ajeet @AjeetThinks]

agent saves you 2 hours, then costs the team 3 in alignment and review. that's negative ROI and no dashboard shows it. i think the next harness problem is the humans around the agent.

[Ajeet @AjeetThinks]

anthropic hit 30,000 merges a day and git tapped out. they built a replacement in 5 months with 7 people. i think most of our dev infra is quietly tuned for humans who commit twice a day, and agents are about to expose it.

[Ajeet @AjeetThinks]

@YashHustle_22 git fetch gives you information. git pull makes a decision. Senior devs prefer knowing what they're merging before changing their working branch.

[Ajeet @AjeetThinks]

An AI agent that solves a task but burns 1,000× the tokens needed is just an inefficient engineer with an API key.

I’m paying more attention to context management, retry loops, and tool calls than another benchmark leaderboard.

[Ajeet @AjeetThinks]

@AnkanXplorer Because storage is cheap and CPU time isn't. Raw formats stay simple and fast to read and write. Compression just trades compute for space after the fact.