Harper is an AI-native commercial insurance company, based in San Francisco and built from scratch. Most knowledge work is judgment locked inside people's heads — the exceptions, the precedents, the decision traces no one ever wrote down. Converting that judgment into software is one of the largest human-to-computational transitions still in front of us, and we think the most honest place to prove it is the hardest one: commercial insurance, a trillion-dollar industry that is still, even now, more than 90% done by hand. We're not patching legacy workflows or adding a copilot to them. We're rebuilding the business so that AI does the work and people do the judgment that AI can't yet — and then teaching it that, too.
It's working: ~1,000 new customers a month and roughly 100x growth in the past year. That pace sets the culture. We're on-site in San Francisco, in the building together, working long days to high standards — because a rebuild this large doesn't happen part-time or by committee. Almost no one joins Harper because they're passionate about insurance. They join because they want to be on the frontier of the AI transition, doing the most consequential work of their career, in a company being built to define a category rather than join one. If that's the work you're looking for, insurance is just where you get to do it.
The role
The companies that win with agents don't win because they found one magic model. They win because they build the harness that turns models into reliable workers: the loop, the tools, the memory, the execution environment, the budget, the fallback path, the guardrails. The model is replaceable. The harness is the company-specific leverage — and at Harper it's the substrate beneath everything, because we run agents on two surfaces at once: product agents that operate the business (intake, sales, service, voice) and coding agents that operate the codebase (frontier coding agents, OpenClaw, our internal Hermes-based agents). Both ride the same harness. You own it. You define the contracts every Harper agent integrates against — the meta-harness wrapping our coding agents and the model-routing layer over our foundation-model providers. Do this right and every new agent ships faster, fails smaller, and learns from the systems that came before it. This is the layer the whole bet runs through.
What you'll own
The agent loop — prompt construction, tool routing, context-window management, retry/timeout/budget logic, multi-step orchestration.
The execution environment — sandbox lifecycle, isolation, blast-radius limits, filesystem and network policy for agents that browse or call APIs.
The tool layer — the canonical set of tools every agent can call. Schema, auth, rate-limit, observability per tool.
Model-provider abstraction — provider routing, fallback chains, cost/latency tradeoffs, eval-driven model selection.
Multi-agent coordination patterns — parent/child handoff, shared-memory contracts, parallel execution, conflict resolution, subagent budgets.
The harness SDK — what every pod engineer imports to ship an agent. If pod engineers ship faster, you did this right.
Guardrails — banned tool combinations, prompt-injection defense, data-egress policy, PII scrubbing.
What we're looking for
8+ years building software, including Senior+ scope at a high-growth company (8–12 years total, 3+ at Senior+).
You've built or owned an agent harness at an AI-substrate company (or a strong open-source equivalent).
You think in tool design, not just prompt design — the prompt is the last 5%, the tool surface is the work — and you can talk trade-offs between agent loops you've used and when each one wins.
You've shipped sandbox infrastructure at scale (Firecracker, gVisor, or comparable isolation).
You write code with AI daily and have strong opinions about which harness behaviors matter and which are theater.
Strong written communication (RFCs, API contracts, integration guides).
Bonus: open-source agent/harness contributions; sandbox/isolation depth; foundation-model partner or early-access experience.
The reality
On-site in San Francisco, in person, long days, high standards. You own the layer every other agent team builds against, so the leverage is enormous and so is the scrutiny — a harness decision you make shows up in every pod's velocity. A rebuild this large doesn't happen part-time or by committee. The right person wants that ownership and that pace.
Logistics
Compensation (OTE): $253,000–$308,000 cash (base + target performance bonus), plus competitive equity.
Location: San Francisco, in-office. Based here or willing to relocate.
Benefits: Uber commuter benefits; breakfast, lunch, and dinner provided; snacks and coffee stocked; free gym membership; health, dental, and vision.
Process: Founder call (15 min) → CTO deep-dive (60 min, harness architecture) → Super Day on-site → founder + CTO offer. No committee. Best offer, first.
To apply: If you've built a harness other engineers couldn't live without — send your resume, the harness, and tell us about a model-routing or sandbox decision you'd defend in five years.