-
-
TYPE OF WORK
Part Time
-
-
SALARY
$1500–$2,000/month
-
-
HOURS PER WEEK
TBD
-
-
DATE UPDATED
Sep 15, 2026
K3Labs.ai builds accounting and operations software — Laoh, a patent-pending accounting engine, and Weaver, a platform on top of it. Small team, real customers, real revenue.
We build LLM-first. Most of our code is written by agents. What we need is someone who can run those agents well enough to ship real features, not demos.
What you'll do:
Take a roadmap feature from description to shipped, working code, using Claude Code and similar agentic tooling as your primary means of building
Run several agents in parallel on separable pieces of work, and keep the pieces coherent when they land
Decompose large features into units an agent can actually complete — this is most of the skill
Validate everything. Agent output that looks right and isn't is the main risk in this way of working, and catching it is your job, not ours.
Work efficiently with context and tokens: scope tasks tightly, avoid re-explaining the codebase every session, know when a fresh context beats a long one
Set up the guardrails — tests, types, review passes — that make agent output trustworthy at speed
Requirements:
Real software engineering ability. You need to read code critically and know when something is wrong. This role is more demanding of judgment than a normal engineering job, not less, because you're reviewing far more code than you write.
Deep hands-on experience with agentic coding tools — Claude Code, Cursor, Codex, or similar. Daily use, on real projects, not experiments.
You've hit the walls already: agents that confidently produce broken code, context that degrades over a long session, refactors that quietly break something three files away. I want to hear how you handle those.
Comfort with ambiguity. You'll get a feature description and a codebase, not a spec.
Overlap with US Pacific time.
Nice to have: TypeScript/Node, Postgres, MCP servers, anything in accounting or fintech, experience with formal testing or property-based testing.
We pay above local market for this role because the skill is rare and we know it. If you're good at this, you'll have other options — what we offer is small-team autonomy, a codebase you'll shape rather than maintain, and a role that grows with the company.
How to apply:
No cover letter. Answer these:
1. What agentic tools do you use, and what does your actual workflow look like? Walk me through taking a mid-sized feature from description to merged.
2. Describe a specific time an agent produced code that looked correct and wasn't. How did you catch it? What changed in your process afterward?
3. Invariants exercise.
Our core product is a ledger — it records financial transactions over time, and every transaction must net to zero across accounts. Write down a minimal set of invariants that must hold for the ledger to be in a valid state.
Then, in your own words:
Why these invariants and not others? What was the design choice at each one?
For each: where would you enforce it — type system, database constraint, runtime check, test suite? What does that choice cost you, and what does it buy?
Which invariant is doing the least work? What would you drop if forced, and what breaks?
Are any of them implied by the others? Show it, or argue why not.
Which of these would an agent be most likely to violate while writing code that passes review? How would you make that violation impossible rather than unlikely?
Where does your set diverge from how accounting actually works in practice — and is that a flaw or a feature?
Use any tools you like, including LLMs — we do. But you will defend this on a call, without notes, and we'll push on the choices. The write-up isn't the test; it's what we'll examine you on.
4. Link to something you've built where most of the code was agent-written.
Shortlisted candidates get a paid trial: one real feature from our backlog, on our codebase, with a call afterward to walk through your decisions.
