Manish Karney

Explorations

Things I’m taking apart, and what they’re teaching me.

So much is suddenly possible, and the boundary is moving faster than anyone’s intuition. I keep my own judgment honest by building at that edge myself: coding agents from scratch, cloud-class models on a laptop, agents loose in messy real systems. Then I watch exactly where they break. These are working notes from the bench — what’s real now, where it still hits a wall, and what I don’t yet know.

Agent systemsStill building

The model is the cheap part

Building a coding agent from first principles — first in Bun, now in Rust — because re-deriving a thing is the only way I really understand it.

Taught meYou learn an agent’s design by trying to rebuild it from scratch.

Still openIf the model is the commodity, how much of an agent’s trustworthiness is just the harness around it?

AgentsTrustSystems
Read the exploration →

Agent limitsOpen question

Confidently wrong is the failure that costs you

Building agents, the dangerous failures were never refusals. They were the confident mistakes — a clean, plausible, wrong action the harness had no reason to stop.

Taught meAn agent’s confidence tells you nothing about whether it’s right.

Still openHow do you build a system that knows when it doesn’t know, when the model itself often can’t tell you?

AgentsTrustCalibration
Read the exploration →

Agents in productionIn production

The boring half of an agent is the half that matters

A personal agent that wakes, reads context, decides, and acts — where the reasoning took a weekend and the trustworthiness took everything after.

Taught meA demo agent needs a good prompt. A real one needs idempotency.

Still openWhat is the least scaffolding that makes an autonomous loop safe to let touch your life?

AgentsReliabilityTrust
Read the exploration →

Model limitsOngoing

Most “reasoning” bugs are memory bugs in disguise

The more agents I build, the more “the model can’t reason” turns out to be “the model never saw the one thing that mattered, at the moment it mattered.”

Taught meMost reasoning failures are really context and retrieval failures wearing a costume.

Still openAs context windows grow, does memory get solved, or just deferred to a bigger version of the same problem?

ContextMemoryAgents
Read the exploration →

EvaluationOngoing

The number went up; the thing got worse

I run batch evals on a voice agent in production. Watching the score climb while real calls degraded taught me more than the score ever did: the eval measured the part that was easy to measure.

Taught meA benchmark you optimize against stops measuring the thing you cared about.

Still openHow do you evaluate a system for the failure you haven’t imagined yet?

EvalsHonestyProduction
Read the exploration →

Real systemsIn production

Reality is the dependency nobody declares

A clean agent meets a legacy desktop app with no API, a form that moved overnight, a PDF that is secretly a photo. The demo was never the hard part.

Taught meEvery system that works in the demo meets a reality it was never told about.

Still openHow much of “AI doesn’t work in production” is the model, and how much is the world being messier than any training set?

ProductionMessinessReliability
Read the exploration →

InterfacesIn progress

The hardest integration is into the system with no door

A voice front-end and a worker that drives a legacy desktop app through computer-use — because the most consequential software has no API, only a login screen.

Taught meWhen there’s no API, the agent has to use the software a human would.

Still openWhen the screen is the only source of truth, how far can an agent be trusted to act on what it sees?

Computer-useLegacy systemsReliability
Read the exploration →

Local AIOngoing

Running the model where the data already lives

Pulling cloud-class models down onto a laptop and phone — a multi-billion-parameter speech model on Apple Silicon — to find where local AI really breaks.

Taught meOn-device isn’t a fallback. It’s the strongest privacy promise you can make.

Still openHow much of the cloud’s job can move on-device before the constraints stop being worth it?

Local AIPrivacyLatency
Read the exploration →

Founding · Healthcare AINow: in pilot

What founding in healthcare is teaching me

I left scale to find out what I’d build with my own name on the mistake. Right now that’s a clinical AI in pilot, and the company is teaching me faster than the code. These are notes from inside it, while I still don’t know how it ends.

Taught meThe further in I get, the more the hard part is who the system is accountable to, not what it can do.

Still openI don’t know yet whether this becomes a company, a product inside someone else’s, or the education that tells me what to build next.

FoundingHealthcare AITrust
Read the exploration →

DirectionOpen question

After scale, what do I want to build?

A decade at Google taught me to build for billions. It didn’t tell me what’s worth building next. That’s the question I’m actually sitting with.

Taught meAfter enough scale, the question stops being “can this reach millions?” and becomes “is this a problem I can respect up close?”

Still openIs the work I want defined by its reach, or by how close it lets me stay to the people it’s for?

DirectionJudgmentScale
Read the exploration →

DirectionOpen question

Joining a lab, or building my own

Two futures I keep weighing: take my judgment into a frontier lab, close to the systems everything is about to be built on, or metabolize it into a company of my own. I haven’t picked, and I want to choose honestly rather than by default.

Taught meThe fork isn’t lab or startup. It’s whose uncertainty I want to absorb every day: the frontier’s, or the customer’s.

Still openWhere does my judgment compound more: inside a mission someone else sets, or one I have to set myself?

DirectionFoundingFrontier AI
Read the exploration →

CraftAlways

Still making things with my own hands

Python services, Swift, Kotlin, this site by hand — staying close to the build so my judgment stays calibrated to what is true right now.

Taught meThe leaders I trust never fully left the keyboard.

Still openWhat does a leader lose the day they stop making things themselves?

CraftLeadershipTaste
Read the exploration →

Personal AI · HousingNow: renovating

Buying a house with AI as my co-pilot

A first-time buyer from a house-poor background, no inherited playbook, against a process built for people who already have advisors. I ran it with AI — modeling, legal, the offer, and now the renovation.

Taught meAI turns an opaque, advisor-gated process into one a careful first-timer can actually run.

Still openWhere does AI level the field for people without generational knowledge — and where does it quietly mislead them?

Personal AIAccessJudgment
Read the exploration →