HardtechOS Collective

Domain-specific agent harness for hardtech startups.Est. 2026, Cambridge, MA.

MissionBuild software & AI platforms for startups creating a more abundant and sustainable world.

Building with Eden Geopower and LithiosEden GeopowerLithios

Disconnected tools make it difficult to use AI.

01 Design
You

I’m designing the next version of an electrochemical cell that pulls lithium out of brine, with two electrodes facing each other. If I tighten the gap between them, what do I gain and what do I risk?

Claude

In general, a tighter gap cuts the cell’s resistance and so its voltage, and raises the risk of shorts, uneven flow, and gas or deposits bridging the gap. I can’t see your drawings or your earlier cells’ test runs, so I can’t tell you where your gap should sit.

01 Design You

I’m designing the next version of an electrochemical cell that pulls lithium out of brine, with two electrodes facing each other. If I tighten the gap between them, what do I gain and what do I risk?

Claude

In general, a tighter gap cuts the cell’s resistance and so its voltage, and raises the risk of shorts, uneven flow, and gas or deposits bridging the gap. I can’t see your drawings or your earlier cells’ test runs, so I can’t tell you where your gap should sit.

02 Operations
You

Our electrochemical extraction unit is running in the field. How do I trace which cell build, and which pre-ship test results, went into each of its stacks?

Claude

Usually you’d keep a record that ties each stack’s serial to its cells’ build and test records, then join the unit’s run data to it. I can’t see your run data or those records, so I can’t tell you which build is in which stack.

02 Operations You

Our electrochemical extraction unit is running in the field. How do I trace which cell build, and which pre-ship test results, went into each of its stacks?

Claude

Usually you’d keep a record that ties each stack’s serial to its cells’ build and test records, then join the unit’s run data to it. I can’t see your run data or those records, so I can’t tell you which build is in which stack.

At the frontier of science & engineering, agents need verified tools & information.

“Harness engineering” is how we connect LLM intelligence with company-specific tools & information.

Claude

What’s the most likely way a chromium electrode feedthrough fails under repeated thermal cycling, and how should we test for it?

Most likely, the scale is what fails: cycling cracks the Cr₂O₃, and contact resistance climbs. A standard thermal-shock test on the bare electrode will show it.I can’t tell which failure mode matters most without your design. If you can share how your feedthrough is clamped and your thermal-cycling results, I can weigh in on what to test.

That’s not how ours are built.

Claude

What’s the most likely way a chromium electrode feedthrough fails under repeated thermal cycling, and how should we test for it?

Most likely, the scale is what fails: cycling cracks the Cr₂O₃, and contact resistance climbs. A standard thermal-shock test on the bare electrode will show it.I can’t tell which failure mode matters most without your design. If you can share how your feedthrough is clamped and your thermal-cycling results, I can weigh in on what to test.

That’s not how ours are built.

Is it all just polished demos & SF Twitter hype?

Zack Kanter
@zackkanter

I finally asked what tool he used most when he sat down to do work. “Python” was his answer, which changed my whole understanding of how they built rockets.

The key reagent for accelerating our ability to dominate the world of atoms is modeling more and more physical industries and domains in code.

Blake Scholl
@bscholl

More than a decade in, the most insane realization is that software is the key to reaccelerating hardware engineering.

Most hardware engineering is actually software engineering in disguise.

Standard Physics
@StandardPhys

We are introducing an AI engineer that turns physics simulations into better designs.

It redesigned a data centre’s cooling system to use less electricity and found a turbine blade that beat a leading published design.

Brian Armstrong
@brian_armstrong

I now have one lead agent for each area of focus, running all the time, and a chief of staff agent that oversees them.

Everything I need to decide or review comes into one prioritized queue.

Posts that inspired how we work. Quoted in part; none is an endorsement.

AI native, company-wide “OS”

1

A harness with your company’s context.

Your objects, connected through one ontology.

2

Tools that map to your ontology.

The agent works through them, record to record.

Supplier
Purchase order
Part
Component
Cell
Simulation
Unit
Field deployment
Brine
Experiment
Test result
Characterization
Agent sees the graph
You

How does lithium recovery compare across our cell materials and brines, and which characterizations explain the difference?

Agent

Material B recovers the most lithium in all three brines. Its lead is smallest in Brine 3. XRD shows Material B keeps its structure over the run; A and C don’t.

Records used: experiments, cells, brines, test results, characterizations. Illustrative.

Agent sees the graph
You

How does lithium recovery compare across our cell materials and brines, and which characterizations explain the difference?

Agent

Material B recovers the most lithium in all three brines. Its lead is smallest in Brine 3. XRD shows Material B keeps its structure over the run; A and C don’t.

Records used: experiments, cells, brines, test results, characterizations. Illustrative.

Mind-blowing AI features are built atop solid data engineering.

1Model

Build the ontology: one model of what the company does and how it all maps together.

Every record has a place, and every link is real.

2Legibility

Make it legible to people and agents: a system of record and a company brain on that ontology.

Knowledge sharing across the org goes up a lot, on its own.

3Automation

Give people and agents tools they can use to execute and accelerate work.

Often where most of the value lands.

4Innovation

Insert intelligence at key moments: choosing the target, designing the experiment.

Closed-loop science.

Mind-blowing AI features are built atop solid data engineering.

1Model

Build the ontology: one model of what the company does and how it all maps together.

Every record has a place, and every link is real.

2Legibility

Make it legible to people and agents: a system of record and a company brain on that ontology.

Knowledge sharing across the org goes up a lot, on its own.

3Automation

Give people and agents tools they can use to execute and accelerate work.

Often where most of the value lands.

4Innovation

Insert intelligence at key moments: choosing the target, designing the experiment.

Closed-loop science.

Model: build the ontology.

We start perfectly unopinionated about the shape of your data model. It’s your software, shaped to fit you.

Four pillars. Every company’s data lives in some of each.

Science

Studies, experiments, samples, results

Hardware

Parts, suppliers, components, units, QA

Field

Deployments, operations, observations

Commercial

Projects, grants, techno-economic models

Legibility: people and agents can see everything.

We increase the connectivity of your data, so everybody can read it.

Scientist“Which characterization explains it?”
Engineer“Which material goes in the next cell build?”

Lithium recovery, by cell material and brine

Brine 1Brine 2Brine 3
Material AMaterial BMaterial C
AgentCan see every record behind it, from the experiment to the ICP run.
Founder“Is Material B the bet?”
Field lead“Which brine is closest to our site’s?”

Automation: tools that do the work.

Start with what people actually spend their time on.

Writing progress reports, and sending them out.Hours a week, gathered by hand from everywhere.
Meetings, Slack, docs and email
AgentDrafts and sends the report
A report, an email, a slide or a dashboard
Opening CAD to understand the as-built.An engineer makes the drawing, or has to inspect one.
CAD models, and the BOM
AgentAnswers from the model
“Which revision of the seal carrier went into Unit 4, and what changed?” Answered from the CAD files and the BOM, with links.
Turning test data into a QA report.Rig and lab results land in a folder.
Rig and lab results
AgentFlags anomalies, drafts the report
A QA write-up, for an engineer to approve

Examples of what we build. Illustrative.

Innovation: intelligence at key moments.

The wheel of science, and where intelligence goes in.

Design Build Test Learn Target

Outside the wheel: what to make, and how good it must be.

Two places for intelligence: choosing the target, and designing the experiment. With all the company’s data, models can reason about what to try next.

Closed-loop science, at its best.

It’s very hard to get there. It’s what we’re striving for, and where we’re headed.

HardtechOS

Hardtech modules.

Four pillars of modules that touch your domain.

Agentic Platform.

The agent, the company brain and the tools that run across every pillar.

Infrastructure.

Foundational to building, hosting, and maintaining an in-house software platform.

All three layers are Collective IP: they belong to the collective.

Collective IP
Hardtech modules the four pillars

Science

Paper repositoryScientific recordsActive learningLab instruments

Hardware R&D

CADParts and procurementSimsHardware recordsAssembly trackingQA rigs

Field

Command and controlWorkcellsFleetDigital twin

Commercial

TEA engineCRMGrant trackingProgress reporting
Agentic Platform across all four pillars
Claude SDK agentFriction reportsOS as MCPCompany brainMeeting transcriptionOS App StoreSlack appsMobile webEdge clientClaude Code skills
Infrastructure
Railway hostingS3 for large filesClerk sign-inRBACGuest accessCI/CDTestingPreview environmentsSecret managementBackupsObservability
Collective IP
Hardtech modules the four pillars

Science

Paper repositoryScientific recordsActive learningLab instruments

Hardware R&D

CADParts and procurementSimsHardware recordsAssembly trackingQA rigs

Field

Command and controlWorkcellsFleetDigital twin

Commercial

TEA engineCRMGrant trackingProgress reporting
Agentic Platform across all four pillars
Claude SDK agentFriction reportsOS as MCPCompany brainMeeting transcriptionOS App StoreSlack appsMobile webEdge clientClaude Code skills
Infrastructure
Railway hostingS3 for large filesClerk sign-inRBACGuest accessCI/CDTestingPreview environmentsSecret managementBackupsObservability

The Collective.

Three companies, one stack. Each has its own implementation on top. The Collective IP underneath is shared.

Eden GeopowerDesign, build, test, learn
LithiosRecord, plan, measure, learn
Client 3Their wordsNew work
Collective IP shared, below every company’s implementation
Hardtech modulesImprovedA hardtech moduleAgentic PlatformInfrastructure
  1. 1

    We build something for Client 3.Work they asked for, in their own implementation.

  2. 2

    It turns out to improve a hardtech module.The collective’s software, below every company’s implementation.

  3. 3

    So Eden and Lithios get it too.And when their work improves a module, Client 3 gets that.

The Collective.

  1. 1

    We build something for Client 3.Work they asked for, in their own implementation.

  2. 2

    It turns out to improve a hardtech module.The collective’s software, below every company’s implementation.

  3. 3

    So Eden and Lithios get it too.And when their work improves a module, Client 3 gets that.

Current clients.

Eden Geopower

Eden Geopower

High voltage pulsed-power for sub-surface resource extraction.

Lithios

Lithios

E-chem process for mineral extraction from brines & wastewater.

We don’t sell SaaS here.

Yours, always

  • The code base
  • Your ontology and your models
  • Your team

You own the code base. You can take it and walk anytime, hire your own software team and fire us.

The collective

  • You benefit from the collective.
  • Your work benefits the others in it.

Making that possible is our investment.

We are members of your team, and we only work with companies we’re inspired by and want to see win.

How we work.

In person

We only work with teams we can work with in person, hands-on with your engineers and founders.

Brian Swartz

Your principal, for now: thought partner and executor.

Development team

We can scale up the effort when you want to.

The collective

The collective supports us: shared Collective IP, and fractional CTO support.

How we work.

Brian Swartz

Your principal, for now: thought partner and executor.

Development team

We can scale up the effort when you want to.

The collective

The collective supports us: shared Collective IP, and fractional CTO support.

Getting started.

In your office

1 full day a week in-office together, side by side with your engineers and team.

Full access

Slack and your company context: Drive, Notion and the rest.

15 hours a week

To start, about 1.5 days a week. Ramp up after 30 days, if you want to.

Tell us what you’re building.

HardtechOS Collective. Est. 2026, Cambridge, MA. Drawings use Hairline (MIT). Type: Saira, IBM Plex Sans, JetBrains Mono (SIL OFL 1.1).