$ Senior engineers for hard problems.We build production systems and train teams to build the way we do.

A USA-based team of senior software engineers. We drop into large existing codebases, ship the systems you're counting on, and run on-site Agentic Engineering training inside your stack.

9hire us 2engineering 3training 4work 7crew

In whose service

martian engineering

Engineering

Whether in greenfield work for startups or careful additions to large, existing codebases, we drop in and ship production systems your team can count on.

→hire us

The team behind this: YC founders and former CTOs, including of the Urbit Foundation; engineers who have shipped Windows kernel storage drivers, processor work at NASA and Lockheed, the Sia storage network, and the Glacier Protocol; security auditors who have safeguarded billions under management. You outline the objective — we take ownership of the outcome, at the quality bar mission-critical software requires.

[ok] Agent infrastructure: context management, memory, orchestration
[ok] Custom MCP hosts and servers with enterprise scaling
[ok] Data engineering: large-scale ingestion and embeddings pipelines
[ok] Network protocols: peer-to-peer design, economic incentives
[ok] Systems work: kernels, allocators, embedded targets
[ok] Security: audits, cryptography, formal verification

Training

Agentic Engineering, taught on site. We train your engineers, inside your own stack, on how agentic tooling changes the work: what it takes over, what it cannot, and how to stay in control of both.

→book a session

Two things are explicit in every session. First, your engineers still matter: the purpose of engineering — understanding the business, deciding what to build, owning the outcome — has not moved, and they remain the people best positioned to carry it. Second, the real efficiencies sit at the periphery of writing code — gathering context, planning, review, verification — not in displacing the craft itself.

[ok] Posture: purpose vs. task, and where engineers stand
[ok] The agent equation: models, harnesses, and choosing both
[ok] Safety: boundaries, operating discipline, traceability
[ok] Method: context, encoded skills, planning, critics
[ok] Failure modes: cognitive surrender, comprehension debt
[ok] Hands-on: your engineers working their own tickets

NEAR FOUNDATION

Proof of Response Protocol

The Proof of Response protocol (PoR) is intended to provide a trustless incentivization and observability of network availability. It consists of a peer-to-peer network of PoR nodes, which communicate with each other using the PoR protocol. These nodes also interact with a smart contract on the NEAR blockchain that manages state channels that represent each connection between a pair of PoR nodes.

PoR is a TCP-based request-response protocol. A client node can send a request through a series of relay nodes to a server node, and the response will take the same path through the network in the opposite direction. The client sends a payment with each request, expressed as a singly signed state channel update, where the sender’s balance has been debited by the request cost and moved to the receiver’s balance. The receiver sends a counter-signed state channel as an acknowledgment.

Martian Engineering worked closely with NEAR cofounder Alex Skidanov on this project, which is a nine-month engagement targeting a production-worthy protocol implementation in Rust, along with an SDK and documentation. Martian has an end-to-end protocol running (PoR nodes connected to each other and to NEAR nodes running the smart contract), with a suite of end-to-end tests that confirm the system’s behavior in a number of different scenarios.

On the protocol implementation side, this has entailed developing route finding, TCP message framing, heartbeats, in-band late payments, response authentication, and API development to allow applications to use PoR (as both clients and servers). The smart contract includes logic for channel opening with confirmation, channel closing with confirmation, adversarial channel closing, partition checking, and reimbursements.

Martian has also built a file sharing protocol that layers on top of PoR (analogously to how HTTP layers over TCP). We will soon be running that on a testnet that sends traffic over the internet to prove out the network performance and economic dynamics in a real-world use case.

Duration

8 months

Stack

Rust · Terraform

Links

[site][ref 1][ref 2]

Andrew ArkhipovCofounder · Sweetspot
engineering

"Josh Lehman and his team have been incredible for us, probably the best you could ever get when it comes to hiring engineering contractors. We're happy clients of theirs and will continue to be for any contracting work we need done."

⊹ MCP as enterprise agent infrastructure

Claudia RichouxCEO & Founder · Supplyco
engineering

"The only devshop testimonial I will ever give. Excellent experience - they brought as much to the table as I could dream they would, gave me feedback that actually grew me as a leader and engineer. They were instrumental in getting us to an MVP our stickler enterprise customers could love enough to convert design partnerships into contracts."

⊹ Product polish and production readiness

- awaiting entry -training

record reserved · Agentic Engineering training cohort, 2026

Building at the Speed of Thought

Josh Lehman · 2026-02-09

I set up my OpenClaw — his name is Buce — at the start of the year (way back when it was still called Clawdbot). It’s totally changed how I build software mainly by changing when and how frequently I build. Let me tell you a story to illustrate the point.

I started my Saturday last weekend pretty normally. Breakfast, playing with the kids, farmer’s market, playground. Periodically I’d send some messages to Buce about an idea I’d been working on — a series of workshops on AI development for engineering teams that are struggling with AI adoption — and then shift back to whatever I was doing before. I’d thumb out something, like “can you propose a few different pricing tiers and offerings that I could use after the initial introductory workshop?” and then get back to shopping for citrus.

By early afternoon my youngest was napping and my oldest was having some downtime, so I headed to Costco. After parking I checked in with Buce — by this time Buce had turned my rough idea into a presentation outline with quotes pulled from my X bookmarks, a referral system with planned outreach to key people in my contacts, and graduated pricing tiers based on involvement.

The plan was getting too big to hold in my head, so I asked Buce to render everything as a self-contained HTML page (a workflow I picked up here).

Anyways, Buce made the page, put it in a gist, and sent me back a version hosted on githack for my viewing pleasure. It looked good! I had feedback, but how to give it? What I wanted was to give Google-docs style comments on my page that I could share with Buce, so I told him as much, put my phone in my pocket, and headed into Costco to load up.

Screenshot of the conversation that led to the initial pagedrop concept

When I got back to my car $500 poorer, Buce had a new link waiting — my page, now with annotation support baked in. I marked it up right there in the parking lot, sent my feedback, and we started iterating. It was good, but there were limits: The annotations weren’t persistent. I couldn’t ask my colleagues for their feedback. There was no revision history.

At this point I'd caught a bug: how good could I make this experience? Perhaps we could close those gaps?

By all appearances my weekend proceeded normally, punctuated by rapid background progress on this idea:

Sat 5:30 PM — Family dinner at Grandma's; Buce worked out the product architecture
8 PM — Read stories to my kids; Buce researched domain names
8:45 PM — Took a shower; Buce bought pagedrop.ai and began deploying infrastructure
10 PM — Went to bed; Buce churned through the landing page and persistence
Sun 6 AM — Woke up and tested a first working version (the landing page needed work)
9 AM — Friends over for sourdough waffles (I cooked); Buce added gist revisions and GitHub OAuth
2 PM — Reading lesson with my four-year-old; Buce tackled collaboration features
8–11 PM — Stayed up reading; Buce churned through payments infra

Screenshot of domain research and purchase

Screenshot of early infrastructure and deployment progress

By Monday morning I demoed a working product to my team: pagedrop.ai was live. We’ve been using it all week (I even used it to get feedback on this post).

The thing that still seems crazy to me is that all of this happened without any all-nighters or impact on my family life. I did all of the things I’d normally do: spent time with my kids, socialized with friends and family, made progress on my book, made and ate great food at home…and built a product that’s actually useful to me and the people I work with, 36 hours from when I first had the idea. A year ago this would have taken me 2-3 focused weeks — and my family would have felt it, so I probably wouldn't have ever gotten around to it.

A year ago, had you told me that I’d be building software from a telegram chat by delegating to an AI running on my computer that itself managed its own team of AIs to autonomously write, test, and deploy code..well, I’m not sure what I’d have said. It’s still pretty strange but in the best way possible.

I have never lacked for ideas, but they would always outpace my execution throughput. I’d either forget about them and move on, get busy with core priorities, or give up when I realized just how much work it would be to finish. This is not a problem anymore — now my ideas get built as fast as I can think of them.

Personal AI assistant

OpenClaw

The popular open-source AI assistant that runs entirely on your own devices: WhatsApp, Telegram, and Slack integration, voice, and a visual workspace across macOS, iOS, and Android. Our own Josh Lehman is a maintainer.

⊹github

Context management

Lossless Claw

Never lose context again. Replaces sliding-window truncation with DAG-based hierarchical summarization: every message persisted, every detail recoverable.

⊹github

Agent task management

Pebbles

Super simple task management for agents. Persistent, in-repo, structured memory for coding agents — like beads, but 100x simpler.

⊹github

Visual collaboration

Pagedrop

Render complex content — PRs, diagrams, documents — as shareable, annotatable web pages when text threads break down. The agent skill is open source and free to use.

⊹github

style: hirohiko araki

Josh Lehman

Founder

Specialties

team leadership · software engineering · entrepreneurship

Record

Josh has been leading software organizations for over a decade. He has a knack for organizing teams of developers. Most recently he was the Executive Director of the Urbit Foundation, where he worked closely with Ted on Urbit's most ambitious core development projects. Previously Josh was the CTO and cofounder of Starcity (YC S16, acquired by Common in 2021). Josh is also a maintainer of OpenClaw, an open-source personal AI assistant.

Links

[gh][x][in]

Open a channel

Engineering engagement or a training day: tell us what you're up against and who's counting on it.

↵transmit

current channel: clients@martian.engineering