Sam Rajaei

Co-founder & CEO @ Binome. Building @ask_mira_ai

[Sam Rajaei @SamanRj]

Stalled deals. Overdue invoices. Missed follow-ups.
Repeated requests nobody tracked.

Most founders I know are still chasing all of it themselves.

That's why I started building Mira.
This is what she looks like today.

[Sam Rajaei @SamanRj]

Founder life: yesterday was one of those workshop days that gets uncomfortable on purpose.

No new feature, no shiny update, just hard questions, harder honesty, and the kind of decisions you can't half-make. Building in the AI space right now means the ground keeps shifting under you, and every few weeks you're forced to ask: are we solving the right problem, for the right people, in the right way?

I don't think founders talk enough about these stretches, the ones with no clean answer, just clarity earned the hard way. Grateful for people in the room who push us toward that clarity instead of comfort.

More to come as it unfolds. This is the part of building that doesn't make the highlight reel, but it's where the real direction gets set.

[Sam Rajaei @SamanRj]

@Zai_org Open weights at this capability level changes the buy-versus-build calculus for a lot of teams. The benchmark score isn't the interesting number.

It's how many companies route around the closed frontier labs entirely once a model this capable is free to fine-tune.

[Sam Rajaei @SamanRj]

@immad @mercury What I'd want to know with Command is how granular the control surface actually is.

Moving money needs a different relationship with oversight than drafting an email. 'Stay in control' is doing a lot of work in that sentence.

[Sam Rajaei @SamanRj]

@AnthropicAI Most 'AI replacing coders' takes miss the domain expertise variable entirely. The shift isn't away from coding skill, it's toward knowing the problem well enough to direct the work. That's a different skill and most orgs haven't trained for it yet.

[Sam Rajaei @SamanRj]

@ClaudeDevs Credentials and observability are the unsexy 80% of this. We learned that the hard way building Mira. The agent doing the task is the easy demo.

The agent being allowed to touch the right systems, and someone being able to see what it did after, is the actual product.

[Sam Rajaei @SamanRj]

@levie The model router point is the one people will skip past. We hit this building Mira: knowing when to lean on the model versus when to lean on the tool you already have is most of the design work. Nobody pays for the model choice. They pay for the routing being right.

[Sam Rajaei @SamanRj]

@mntruell @SpaceX Congrats. The interesting part of this deal isn't the model, it's the distribution. Cursor already had developers living inside the tool for hours a day. SpaceX just bought attention, not capability.

[Sam Rajaei @SamanRj]

@McKinsey Data quality is the safe answer. The harder problem is getting it to the right place, in the right context, at the right time. Most companies aren't losing to bad data. They're losing to inaccessible data.

[Sam Rajaei @SamanRj]

@nvidia The energy constraint keeps getting framed as a bottleneck. I read it more as a forcing function. Every time a layer gets constrained, the one above it gets more efficient. The best software usually shows up right after someone says the hardware can't keep up.

[Sam Rajaei @SamanRj]

We built Vera for agents to use' is a bigger shift than it sounds. When the unit of computation changes from human request to agent task, the memory architecture changes with it. Long-context management, parallel tool calls, persistent state; these hit CPUs differently. The fact that hardware is now being designed around the agent model rather than retrofitted tells you where the industry actually thinks this is heading.

[Sam Rajaei @SamanRj]

@OpenAI The shift this points to: from agents as interactive tools to agents as background processes. Save the window, run it on the job that takes 4 hours, not the one that takes 4 seconds. That async-first pattern is where operations-heavy use cases really open up.

[Sam Rajaei @SamanRj]

@pierreeliottlal @ycombinator These are right for 0 to 10. What the playbook doesn't cover: somewhere around customer 200 you're still the person holding everything together. The rules get you to PMF. They don't get you out of being the ops glue. That's the problem we're building Mira around.

[Sam Rajaei @SamanRj]

@pmddomingos Same problem one layer down. Most businesses deploying agents are still figuring out how to align the agent with what they actually care about. Not the safety question. The 'does this thing know what we're optimizing for' question. That's the constraint most operators hit first.

[Sam Rajaei @SamanRj]

Spent the first few days running it on decision loops, not demos. The part that stood out wasn't output quality, it was how it handles the underspecified middle of a task. Fewer 'clarify this before I proceed' exits. More completing the thing. That changes how you architect the workflow.

[Sam Rajaei @SamanRj]

The loop structure is what's interesting. Discovery - implementation - validation, run autonomously over extended periods. Most business process automation fails exactly here: the loop either doesn't close or requires human bailout at every step. What failure modes are you hitting most?

[Sam Rajaei @SamanRj]

@OfficialLoganK 100%. And there's a version of this for scaling companies: great products get undone by the ops that can't keep up with them. At some point the product quality question and the ops quality question are the same question.

[Sam Rajaei @SamanRj]

@thsottiaux More important UX decision than it looks. If you're running longer tasks, knowing exactly when your window opens changes how you sequence work. Most people think of rate limits as a throttle. It's really a scheduling problem. Good to see this treated as one.

[Sam Rajaei @SamanRj]

@garrytan We kept hitting this building Mira. Every tool made us faster at doing things we shouldn't have been doing at all. The cage gets rebuilt in an afternoon. The unlock isn't speed, it's removing the category of work that shouldn't touch you.

[Sam Rajaei @SamanRj]

Using it on decision loops we run internally at Mira. Testing how it handles ambiguous handoff points, when does the agent complete vs. escalate? First weekend of real production-like conditions. Interesting results so far.Using it on decision loops we run internally at Mira. Testing how it handles ambiguous handoff points, when does the agent complete vs. escalate? First weekend of real production-like conditions. Interesting results so far.