Harness Engineer (AI Agent Systems)
Truewind
Software Engineering, Data Science
San Francisco, CA, USA
USD 180k-200k / year + Equity
Posted on Mar 31, 2026
Truewind
AI-Powered Accounting
Harness Engineer (AI Agent Systems)
$180K - $200K•0.10% - 0.25%•San Francisco, CA, US
Job type
Full-time Role
Engineering, Full stack Experience
3+ years Visa
US citizen/visa only Apply to Truewind and hundreds of other fast-growing YC startups with a single profile.
Apply to role › About the role
What this role is
We build AI agents that do real work.
Not assistants. Not demos.
Agents that execute workflows end-to-end and produce correct outcomes.
Your job is to:
- build those agents
- and build the systems that make them reliable
This is not prompt engineering.
This is making AI work in production.
What you’ll do
- Build agents that execute multi-step workflows
- Design systems for validation, retry, and failure handling
- Define constraints (schemas, invariants, contracts)
- Add feedback loops (detect → debug → improve)
- Turn failures into reusable systems
What this role is NOT
- Not prompt engineering
- Not one-shot demos
- Not feature-heavy product work
You are building agents that do the work, and the systems that ensure they do it correctly.
Note: This is different from “vibe coding.” You won’t just prompt and accept outputs. You’ll build systems so results are reliable and repeatable.
What we’re looking for
- Strong systems thinking
- Background in:
- infrastructure, backend, or data systems
- developer tools or internal platforms
- Experience building reliable systems (not just features)
- Comfortable debugging complex, ambiguous problems
Important:
LLM experience alone is not enough.
We care about how you make systems reliable.
Good fit if you:
- Think in constraints, invariants, and feedback loops
- Care about correctness, not just output quality
- Have automated real workflows end-to-end
- Prefer building systems over features
Not a fit if you:
- Mostly prompt models and accept outputs
- Have only built demos or prototypes
- Avoid debugging or failure handling
Application (required)
1. Project (GitHub)
An agent system that:
- performs a multi-step task
- includes validation
- handles failures (retry, fallback, etc.)
2. Short answer (5–10 sentences)
Describe a system where an AI agent failed.
What caused it, and how would you fix it?
How we measure success
- Agents complete real workflows with minimal human input
- Outputs are correct by construction
- Failures decrease over time
- New capabilities come from improving the system, not patching outputs
Why this matters
AI models are already powerful.
The bottleneck is making them:
- reliable
- structured
- production-ready
The teams that win will not have better prompts.
They will have systems where agents actually work.
Before you apply
Most engineers won’t enjoy this role.
It requires:
- thinking in systems instead of code
- caring about correctness instead of speed
- debugging behavior instead of writing features
But if this clicks for you,
you’ll be working on the actual frontier of software engineering.
About Truewind
At Truewind, our vision is a world where more startups succeed, where founders can access more variety of capital options, where every leader in an organization can access financial insights, and where investors can scale to invest in all the best companies.
Truewind modernizes the financial back-office for startups. Through a combination of intelligent software and concierge service, we deliver a delightful financial back office experience. Truewind is doing for the back-office what AWS did for computing — making it easy, scalable, and on-demand.
Truewind is a Y-Combinator company and backed by a set of fantastic set of firms, angels and our customers, including Scrum Ventures, Beat Ventures, Soma Capital. We are innovating on the financial infrastructure for private markets, building the repository for financial performance and reducing the friction to information sharing.