Agent engineering lab · New York & Bangalore
LatentFlux designs, evaluates and ships production agents for teams with rules to follow. Research-grade harnesses, small models that run them cheaply, and a verified record of every outcome.
Trace
Every tool call, prompt and decision recorded
Gate
Spend caps and human sign-off where it matters
Verify
Outcomes checked before anyone gets paid
Run summary
Model-agnostic. Built on the stack you already run.
The LatentFlux approach
A demo agent takes a weekend. A production agent takes discipline. We arrive with the engineering, the evaluation and the controls already worked out.
01 / Evaluation
We replay real runs in simulation before an agent touches production, so a regression shows up as a diff, not a support ticket.
Agents break in the lab, not in front of customers.
02 / Harness
Tools, memory, planning and recovery are engineered as one system. Most agent failures live there, so that is where we spend our time.
Start from proven scaffolding, not a blank page.
03 / Small models
Where a tuned small model keeps up on your task, we use it. Where it doesn't, we tell you and route to the frontier.
Pay for capability you actually use.
04 / Control
Spend caps, approvals, scoped permissions and an audit trail are part of the build, not a later phase. Your data stays in your cloud.
Built for teams that answer to auditors.
How we work
LatentFlux pairs engineers who have shipped agents in regulated settings with an evidence layer that records what the agent did and proves whether it worked.
Forward-deployed engineers
Our engineers work inside your architecture, your rules and your risk. They decide what an agent may do alone, what waits for a person, and what counts as done.
LatentFlux evidence layer
Every run is traced, checked against an agreed definition of done, and signed. Your team, your auditors and your vendors read the same record.
Capabilities
Each one started as a research program in our lab. Each one now ships inside the agents we build, and each one is measured.
Most agent failures we see aren’t the model’s fault. They come from the code around it: how tools are described, what gets remembered, when it retries, how it recovers. We treat that scaffolding as the thing worth studying. Right now that means rebuilding a published agent-memory paper from scratch to see which of its gains survive.
Which harness choices change success rates, and which only change the bill?
A small model tuned for one job can often keep up with a frontier model on that job, at a fraction of the cost. We’re mapping where that holds.
Where’s the crossover?
Agents should break in a simulator before they break in production. We replay real runs so a regression shows up as a diff, not a support ticket.
What would a flight simulator for agents measure?
An agent that reads email will, sooner or later, read one written to manipulate it. We study injection, leaks, and how to scope what an agent may do with what it reads.
Whose instructions should an agent follow?
Some teams can’t send their data to someone else’s model, full stop. We work on agents that live entirely inside an organisation’s own walls.
What’s the smallest stack that’s still useful?
Ad buying and banking have rules for good reasons. Agents working there need spend caps, sign-offs and an audit trail long before they need a personality.
What can an agent propose, and what should always wait for a person?
Built for production
Traceability
Tool calls, model outputs and human approvals are captured as the work happens. Nothing is pieced together afterwards.
Each dot is a step. The odd one out is a human sign-off.
Control
Agents propose; policy decides what runs alone and what waits for approval. Spend caps and scopes are enforced, not suggested.
Steps before and after a human gate.
Efficiency
We benchmark small models against frontier ones on your actual task, then route each step to the cheapest model that holds quality.
Illustrative cost per run: frontier models versus a tuned small model.
Services
Assess
Build
Verify
Run
Built for industries where an agent’s mistake has a cost.
Now taking design partners for 2027
Bring one process that has rules, volume and a cost when it goes wrong. In a working session we’ll map it, scope the first agent and define how we’d prove it works, before you commit to anything.