Karthik Rao

software engineer @oqoqo_ai, ex @microsoft

[Karthik Rao @raokarthik99]

Is it better than using Codex/ChatGPT Work? I personally think grok 4.5 is a super fast model that I want it to be able to use it the same way I use codex with all integrations and computer use. I feel OpenAI is really good at getting computer use right and I felt atleast cursor was a bit sluggish when it came to this functionality. But been hearing grok bot is designed to be explicitly good and fast at this. Excited to try it if yes and hoping the same improvements landing for coding in cursor!

[Karthik Rao @raokarthik99]

I kind of agree with these interfaces being the convergence point.

Let’s predict that all regular and especially white collar people will just be running agents like all developers already are today.

Can we conclude that traditional dashboards and apps are not going to be relevant? Then what’s actually relevant and to whom? Will all apps just become slack bots? Won’t that lead to same fatigue people have today with the million apps?

Personally I like having different apps that do different things even though I mostly will be using something like codex for all work. I still like the idea there’s an app for the product I can use without having to use an agent and ofc. a well designed GUI induces a psychological trust with the company.

It’s just very confusing right now to know what to focus on when building literally anything!

[Karthik Rao @raokarthik99]

and we are just getting started.

if you’re building agents or agent-first products, we want oqoqo to be the easiest, fastest and cheapest way to run realistic evals, custom benchmarks, compare models, and actually see where things break.

we handle the painful infrastructure bits — scaling, concurrency, durable orchestration, new agents/models, integrations (incl DCR/CIMD oauth in mcp), and a shared workspace for your team.

near term:
- more agents (devin, grok build, antigravity, factory, warp…) + any custom ones
- schedules, slack notifs, pr triggers
- pause/resume/retry without losing progress
- better rubrics + code-based criteria
- “deep” mode with adversarial agents
- visual evals on browser/computer use
- public gallery for results

longer term stuff that actually excites us:
- orchestrator agent that deals with race conditions + dag queuing
- simulated world models so we don’t burn real resources
- multi-step evals with both deterministic steps and contextual user simulation

super excited for what’s next.

[Haritha Nair @margharithaa]

@oqoqo_ai is live!

Public benchmarks show how models perform on average, and the results are meant for labs and research teams.

As agents become primary consumers of software, evals and benchmarks need to escape the sandbox and reflect our real world better.

Oqoqo is the easiest way to build evals and custom benchmarks for real world tasks at frontier scale.

Try it for free at https://t.co/akEJoFCUCj

[Karthik Rao @raokarthik99]

I personally think babysitting is quite useful in terms of actually caring and building the right thing. Let’s say somebody promises you that you never to look at what the agent did anymore, would you blindly believe it or would you care about your product enough that you’ll always check and correct? Efficiency gains are already crazy, so there’s quite the ironic boosted value in still steering agents and building the best thing you can

[Karthik Rao @raokarthik99]

@margharithaa so apparently they have rearch. the way they index for the first time since ages. wait for the sep update

[Karthik Rao @raokarthik99]

it’s worse when they write code that start doing heavy operations in memory with no consideration of the underlying infra or the connected arch whatsoever. just as an example try prompting any agent to implement bulk delete of items in a table that involves cascaded deletes and how to keep track of status and update the clients in a durable and reliably manner

[Karthik Rao @raokarthik99]

@almonk It doesn’t help when there are so many apps that to do the same thing. I see so many remote/cloud coding agent apps and so many iMessage based personal agents and I am not sure what problem they are trying to solve that existing solutions don’t solve for.

[Karthik Rao @raokarthik99]

true that. sometimes i'm perplexed by how engineers are also a broad spectrum and act much like consumers when it comes to setup and ease of use. for example, technically you can use any non-gpt/claude model with codex or claude code (if you just read the docs, sigh) but there's still so many coding harnesses with nice ui which people end up flocking to

[Karthik Rao @raokarthik99]

it's because us engineers love evolving our "software brain" which models real world such that we think everything can be solved by just "packaging" or "marketing" the right way. the real world has lots of entropy. humans are weird and don't do "rational" things every single second of the day.

[Karthik Rao @raokarthik99]

The whole success of software industry was all about leverage and avoid reinventing the wheel. People love to say oh that framework is so old school or I just killed this framework and so on. But the reality is imagine that solo developer who shipped something crazy 20 years ago and is the foundation of today’s world. That person is more valuable than any amount of tokens. I hate the culture of brute forcing your way to burn tokens just to sound woke

[Karthik Rao @raokarthik99]

@nateberkopec There’s no validation of anything. No talking to real people. Just random building and calling yourself a hero because you “shipped” a bunch of markdown files. And GitHub stars make it worse by seeming like a project is high quality because it has so many stars. Sigh.

[Karthik Rao @raokarthik99]

@mynameisyahia Silly question but what differentiates a live web extractor from a search product like Exa or Parallel? Is https://t.co/0aEff6MScW for a different purpose or better for the same use cases as these other ones?

[Karthik Rao @raokarthik99]

Totally. It’s very tempting for us engineers to be like oh I should just make this specific need a “platform” and ship an open source SDK. But there’s no validations of what already exists in the market, no talking to people, nothing. Just an ever growing number of libraries which would never get the maintenance and drop off in just a couple of months or so.

[Karthik Rao @raokarthik99]

pro tip: if you are using different harnesses/models for different use cases, just ask codex to delegate using it's CLI instead of trying to find the perfect super A2A or multi agent app

[Karthik Rao @raokarthik99]

til something crazy: not only can you "@ mention" a different thread/chat but you can also ask codex to coordinate with them and wire up cross-thread orchestration and sequential workflows super easily!

[Karthik Rao @raokarthik99]

i want it in the cloud so that it's always on and accessible from anywhere but just something that I control lifecycle and persists all my state/data and secured under machine login that I can RDP/SSH into just like local. I'll set it up once exactly like local and use it for all PRs.

products like https://t.co/dehzkBp58K already give you like this persisted 4vCPUs, 16GB RAM and 128GB disk for $100/m

[Karthik Rao @raokarthik99]

@BenceRedmond i would be more comfortable with an always on remote mac that i setup once and ssh into and use plain old worktrees to parallelize stuff

[Karthik Rao @raokarthik99]

have been trying to get cursor cloud agents to work properly across my SDLC but being able to run the app just like in local with all required secrets/user credentials/saved remote browser credentials/all local services like postgres docker etc. and reusable local data is too much setup work