← harvey / Staff Product Manager, Agent Platform
brief / art_xli0_S1-k9g
role
model
anthropic/claude-sonnet-4.6
created
2026-05-21T22:43
Company snapshot
Harvey is an AI-first legal technology company building end-to-end agentic workflows for law firms and in-house legal teams — not document search or simple Q&A, but agents that pull precedent, draft, analyze, and enforce firm-specific rules autonomously. As of early 2025 the company reports 1,000+ customers across 60+ countries and has raised significant venture backing (reported rounds from OpenAI Startup Fund, Sequoia, and others — exact figures not independently verified here). Harvey's engineering reputation centers on frontier model fine-tuning for legal domain specificity and enterprise-grade compliance infrastructure (ethical walls, audit trails, matter-level governance). The company is in active category-creation mode, competing with emerging legal AI players while differentiating on depth of legal domain integration. Specific internal projects, named executives beyond public record, and precise headcount are not confirmed here.
Team stack
Based on the JD and public signals: Python-heavy backend (likely FastAPI or similar), LLM orchestration layer (likely custom agent runtime, possibly LangGraph or proprietary — JD language suggests proprietary orchestration), vector retrieval for precedent search (likely pgvector or Pinecone — uncertain), React/TypeScript frontend (inferred from standard enterprise SaaS patterns), enterprise auth and RBAC infrastructure for ethical wall enforcement (specific vendor uncertain). The JD explicitly calls out governance, audit trails, and multi-stakeholder approval flows as core product surfaces, suggesting significant investment in compliance infrastructure. Cloud provider likely AWS or GCP (uncertain). Agent platform likely uses structured tool-calling patterns over frontier models (GPT-4/Claude/proprietary fine-tunes — based on Harvey's public positioning).
Likely questions (10)
| area | question | why |
|---|---|---|
| system_design | Harvey's agent platform needs to enforce ethical walls — a lawyer on one matter cannot see documents or context from a conflicting matter. Walk me through how you would design the product and data architecture to make this a first-class constraint, not a bolt-on. | The JD explicitly states 'ethical walls, audit trails, governance, multi-stakeholder approval flows — these aren't afterthoughts here, they're the product.' This is the single highest-signal differentiator from generic AI PM roles. |
| system_design | Describe how you would design an agent creation and discovery surface for a large law firm. How do lawyers find the right agent for a task? How do firm admins govern which agents are available to which practice groups? | The JD calls out 'how agents get created and discovered' as a potential core scope area for this role. |
| domain | You've built multi-agent orchestration systems (OpenClaw). How does designing an orchestration layer for a consumer/SMB use case differ from designing one for enterprise legal, where every agent action may touch a real client matter and carry liability? | JD requires comfort with enterprise governance and the candidate has direct multi-agent orchestration experience — interviewers will probe whether that experience translates to regulated, high-stakes environments. |
| behavioral | Tell me about a 0-to-1 product you owned where the problem definition was as important as the solution. What did you get wrong initially, and how did you course-correct? | JD explicitly states '0-to-1 product work in a space where the patterns haven't been written yet' and lists '0-to-1 or major pivot' as a required credential. |
| coding | You don't need to write code, but walk me through a technical architecture decision you made on one of your platforms — for example, why you chose SSE over WebSockets for streaming RL metrics, or how you structured the RAG retrieval pipeline in Fintellect. What tradeoffs did you weigh? | JD requires 'comfort with technical depth — engage with engineers on system design, architecture tradeoffs, and what's actually hard.' Harvey will probe whether PM depth is real or performative. |
| domain | AI agents in legal produce outputs that go to courts and counterparties. How do you build a verification and trust layer into the UX so lawyers adopt the agent's output confidently without blindly rubber-stamping it? | JD states 'getting this right means building trust systems, verification flows, and governance infrastructure alongside the core experience' — this is a core product design challenge Harvey is actively solving. |
| behavioral | Describe a time you had to make a hard prioritization call with imperfect data under competitive pressure. What framework did you use, and what would you do differently? | JD explicitly calls out 'hard prioritization calls with imperfect data in a fast-moving competitive landscape' as a required capability. |
| culture | Harvey describes itself as moving fast, operating with intensity, and expecting PMs to be hands-on with details while holding a clear vision. Give me a concrete example of a week where you were simultaneously in the weeds on a technical detail and driving a strategic decision. | JD culture language ('move fast, ship with design partners, hands-on with details') signals they will screen hard for PMs who delegate too early or stay too abstract. |
| behavioral | You've led developer platform work at Intuit at significant scale (675M+ engagements, 50K TPS). How do you apply platform-thinking — SDKs, self-service, governance — to an agent platform where the 'developers' are lawyers, not engineers? | The role is explicitly a platform PM role; Intuit experience is the strongest analog on the resume and interviewers will want to see the transfer of platform intuition to a non-technical user base. |
| domain | Walk me through how you would instrument an AI agent platform for a law firm to produce an audit trail that satisfies both internal risk/compliance teams and external bar association obligations. What events do you log, how do you surface them, and who owns them? | JD lists audit trails as a core product surface, not a compliance checkbox. This tests whether the candidate treats governance as a product design problem. |
Talking points
- OpenClaw multi-agent orchestration (StreamIO): Built a production multi-agent gateway with subagent delegation, profile management, and session switching — directly analogous to Harvey's agent creation/discovery surface. Can speak concretely to the design decisions around routing, context isolation, and agent lifecycle management, and can bridge that to the enterprise governance layer Harvey needs (ethical walls, matter-level scoping).
- ICE Self-Service platform at Intuit (developer platform at scale): Reduced developer onboarding from 2–3 weeks to minutes, scaled to 675M+ engagements and 50K TPS — demonstrating the ability to own a platform product end-to-end, instrument it with telemetry, and drive adoption across a large, heterogeneous user base. The self-service and governance patterns (GitOps config, DevPortal, drift detection) map directly to Harvey's agent governance and firm-level deployment constraints.
- RL Workbench + aeval: Built a full post-training evaluation stack (12 RL algorithms, live SSE metric streaming, adversarial safety testing, bootstrap confidence intervals, automated safety gates) — demonstrates the technical depth to engage credibly with Harvey's applied ML and engineering teams on model behavior, evaluation rigor, and trust infrastructure, not just UX.
- 0-to-1 founder experience (Streamio AI + Fintellect AI): Shipped production applications from zero across regulated-adjacent domains (real estate, financial advisory) with RAG pipelines, multi-provider LLM orchestration, and enterprise auth/payments — evidence of owning the full product lifecycle including customer discovery, go-to-market, and iterative refinement under resource constraints.
- NeurIPS 2014 + BRAIN platform: Published researcher in neural networks with a 2026 production rewrite spanning 413 to 8B parameters — establishes credibility as a PM who can engage with frontier AI research teams at Harvey's level of technical ambition, not just translate requirements.