← openai / Product Manager, API Agents
interviewer_questions / art_xl2WIChAWzk
Interviewer
The interviewer profile provided is not a personal LinkedIn profile but rather a mirrored copy of the role's known interview question bank for the PM, API Agents position at OpenAI. No individual interviewer name, tenure, current title, or personal background can be extracted from the input. Based on the question set provided, the interviewer appears to be a technically deep PM or engineering leader who probes system design, SDK interface design, developer empathy, and product philosophy around agentic infrastructure. The questions span both hard technical depth (rate limiting, state management, streaming SDK design) and behavioral/cultural dimensions (backward compatibility, human-in-the-loop tradeoffs). Expect this interviewer to push hard on the candidate's ability to reason at the API primitive level, not just the product strategy level.
My profile through their lens
From the perspective implied by this question set, Felix is an unusually strong candidate because he has actually built multi-agent orchestration (OpenClaw), implemented 12 RL algorithms in a post-training workbench, shipped developer SDKs and platform infrastructure at Intuit scale (675M+ engagements, 50K TPS), and published at NeurIPS — a rare combination of hands-on AI engineering and enterprise platform PM experience. The interviewer will likely be most impressed by the OpenClaw gateway/subagent delegation architecture and the ICE platform work, as these directly map to the role's core deliverables. The RL workbench (GRPO/DPO/PPO across TRL, VeRL, OpenRLHF, NeMo RL) gives Felix peer-level credibility on post-training that most PM candidates cannot claim. The primary risk the interviewer will probe is whether Felix's recent experience is founder/solo-builder depth versus cross-functional platform PM depth at a large organization with competing stakeholders.
Questions they may ask (20)
| category | question | why | how to prepare |
|---|---|---|---|
| resume_deep_dive | Walk me through the OpenClaw multi-agent orchestration framework — specifically the gateway protocol design. What primitives did you expose, how did you handle subagent delegation failures, and what would you change if you were building it for 10,000 external developers instead of your own product? | OpenClaw is the most direct analog to the role's core deliverable. The interviewer's system design question #1 about durable, resumable multi-step agent execution maps exactly to what Felix built. The follow-up about external developers tests whether his thinking scales from founder-mode to platform-mode. | Prepare a crisp 3-minute architecture walkthrough of OpenClaw: gateway protocol, subagent delegation model, session/state management, and failure handling. Then explicitly articulate what you'd change for a public API — versioning, error taxonomy, observability hooks, and rate limiting at the fan-out layer. |
| resume_deep_dive | At Intuit you scaled ICE from 6K to 50K TPS via rSocket migration supporting ~1.5M concurrent connections. What were the hardest product decisions you made during that migration, and how did you balance developer backward compatibility against the performance gains? | Interview question #9 directly asks about backward compatibility vs. research/infrastructure velocity. Felix's ICE migration is the strongest concrete evidence he has navigated this exact tension at scale. The interviewer will want to hear the product decision-making, not just the technical outcome. | Prepare a specific story: what the breaking change was, how you communicated it to developer stakeholders, what the migration path looked like, and what you'd do differently. Quantify the developer impact (onboarding time reduction, TPS gains) alongside the compatibility cost. |
| resume_deep_dive | Your RL workbench benchmarks GRPO, DPO, PPO, and 9 other algorithms across TRL, VeRL, OpenRLHF, and NeMo RL. As a PM, how did you decide which algorithms and frameworks to include, and how would you translate that benchmarking methodology into a developer-facing product decision at OpenAI? | OpenAI's rapid model cadence (GPT-5.3, 5.4, 5.5) means the API Agents PM must constantly make decisions about which capabilities to surface and when. Felix's RL workbench shows he can reason about algorithm tradeoffs — the interviewer will probe whether that translates to product prioritization instincts. | Frame your algorithm selection as a product prioritization exercise: what signals (developer demand, capability gaps, safety considerations) drove inclusion. Then bridge to how you'd apply that same framework to deciding which agentic primitives to expose in OpenAI's API roadmap. |
| resume_deep_dive | You reduced developer onboarding at Intuit from 2-3 weeks to minutes via the ICE Self-Service platform. Walk me through how you identified that onboarding friction was the right problem to solve, and how you measured success beyond the time-to-production metric. | The role requires deep developer empathy and the ability to identify high-leverage problems. This question tests whether Felix's onboarding win was insight-driven or execution-driven, and whether he thinks in terms of developer outcomes beyond vanity metrics. | Prepare the discovery story: what data (SQL/BigQuery usage analysis) and qualitative signals (developer interviews) led you to prioritize onboarding. Include secondary metrics — adoption rates, support ticket reduction, developer satisfaction — and be honest about what you didn't measure. |
| technical_domain | Design a durable, resumable multi-step agent execution API for OpenAI's platform. What primitives would you expose — runs, steps, checkpoints, tool_calls? How do you handle partial failures mid-execution, and how do you think about state management when the agent has already consumed tokens and made external side effects? | This is system design question #1 from the known question bank. Felix's OpenClaw architecture and his Intuit platform work are directly relevant, but the interviewer will push on the external API design layer, not just internal implementation. | Sketch the API surface: Run object (id, status, created_at, checkpoint_id), Step events (tool_call, tool_result, model_turn), and a resume endpoint. Address idempotency keys for tool calls with side effects, checkpoint granularity tradeoffs, and how you'd expose this in the SDK with streaming events. |
| technical_domain | Sketch the interface design for a streaming agent run events SDK method. What events do you expose — thinking, tool_call, tool_result, message_delta, error, done? How do you handle backpressure when the consumer is slower than the model, and what's your error taxonomy? | This is coding/SDK question #3 verbatim from the question bank. Felix built real-time HLS streaming with WebSocket communication and SSE metric streaming in his RL workbench — he has direct implementation experience to draw from. | Design the event schema explicitly: event type enum, payload structure, sequence numbers for ordering. For backpressure, discuss buffering strategies, client-side flow control, and what happens on buffer overflow. Error taxonomy: distinguish transient (rate_limit, timeout) from fatal (context_exceeded, tool_error, safety_block) errors. |
| technical_domain | How do you think about the tradeoff between flexible low-level primitives versus opinionated higher-level abstractions in an agents SDK? Give me a concrete example from your own work where you made this call, and tell me where you think OpenAI's current Agents SDK sits on that spectrum. | This is domain question #5 from the question bank. Felix has built both low-level infrastructure (rSocket migration, GraphQL API for Asterias) and higher-level abstractions (OpenClaw orchestration, ICE Self-Service). The interviewer wants to see if he has a principled framework, not just opinions. | Use OpenClaw as your primary example — explain what you made opinionated (subagent delegation protocol) vs. flexible (tool integration interface). Then give a genuine, informed take on OpenAI's Agents SDK: where the abstraction level is right, where it's too opinionated, and what you'd change. |
| technical_domain | What are the top 3 developer pain points building production agentic apps today, and what specific API or product change would most directly address each? | This is domain question #4 verbatim. Felix has built production agentic systems (OpenClaw, Fintellect Agents, AutoEval) and has customer discovery experience from both Streamio and Fintellect. The interviewer will probe whether his pain points are grounded in real developer experience or theoretical. | Anchor each pain point in a specific experience: (1) non-deterministic failure recovery — from OpenClaw subagent delegation failures; (2) observability/debugging in multi-step runs — from your RL workbench SSE streaming work; (3) cost/latency unpredictability in fan-out — from your Intuit TPS scaling work. Pair each with a concrete API primitive or product change. |
| gap_transition | You've been running two founder-mode startups since September 2024. At Intuit and Splunk you operated in large cross-functional orgs with competing stakeholders, engineering partners, and GTM teams. How do you think about re-entering that environment, and what specifically do you think you'll find harder than your founder experience? | The 9-month founder gap is the most obvious transition risk. The role requires driving consensus across research, engineering, and GTM at a company with hundreds of engineers — a very different motion from solo/small-team building. | Be honest and specific: name what you'll miss about founder mode (speed, full context) and what you're genuinely excited to re-enter (scale, research collaboration, cross-functional depth). Avoid generic 'I thrive in both' answers — show self-awareness about the adjustment. |
| gap_transition | Your most recent large-org PM role was at Intuit, ending September 2024. OpenAI's API Agents team moves at a pace where the model roadmap can invalidate your product decisions in weeks. How do you think your Intuit experience — where you were working with more stable infrastructure — prepares you for that velocity? | Intuit's platform work was high-scale but relatively stable infrastructure. OpenAI's GPT-5.3/5.4/5.5 cadence and the rapid evolution of agentic capabilities represent a fundamentally different product environment. The interviewer will probe whether Felix can operate in that ambiguity. | Point to your RL workbench work (benchmarking across 4 rapidly evolving frameworks) and your Splunk Scheduler delivery (4 months, end-to-end) as evidence of operating in fast-moving technical environments. Acknowledge the difference in org scale and frame it as a strength — you've seen both sides. |
| gap_transition | Your AI/ML work — RL workbench, aeval, BRAIN rewrite — appears to be largely self-directed research and solo builds. How do you translate that depth into influencing a research team of PhDs at OpenAI who are making post-training decisions that affect your product roadmap? | The role requires partnering with research teams 'at a technical level.' Felix's NeurIPS publication and RL workbench give him credibility, but the interviewer will probe whether he can operate as a peer-influencer in a research org versus a solo builder. | Reference your NeurIPS publication and your RL workbench (GRPO/DPO benchmarking) as evidence of research-level thinking. Then give a specific example from Intuit of influencing engineering or research decisions through data — your Service Language Assessment presented to the CTO is a strong anchor. |
| behavioral_situational | Tell me about a time you drove alignment across research, engineering, and GTM on a technically ambiguous product decision. What was the decision, who disagreed, and how did you resolve it? | This is behavioral question #6 verbatim. Felix's Intuit work (Service Language Assessment to CTO, MSaaS Drift Detection program, GCP-to-AWS migration) has multiple candidates for this story. The interviewer wants to see structured conflict resolution, not just consensus-building. | Use the Service Language Assessment as your primary story: 9 languages, competing engineering factions, CTO-level stakes. Structure it as: what the ambiguity was, who the stakeholders were, what data you used to build the case, what the dissenting view was, and how you drove to a decision. Be specific about what you gave up to get alignment. |
| behavioral_situational | Describe a developer platform feature you shipped that turned out to be the wrong abstraction. What signals told you, and what did you do? | This is behavioral question #7 verbatim. This is a high-risk question because it requires genuine intellectual honesty about a failure. Felix's extensive platform work (ICE, SDK Starter Kits, Asterias GraphQL API) gives him material to draw from, but he needs a real example, not a sanitized one. | Identify a real wrong-abstraction story from your Intuit or Splunk work — ideally something where developer adoption data (BigQuery/SQL) told you the abstraction was wrong before stakeholders admitted it. The Asterias declarative asset lifecycle platform or the SDK Starter Kit scaffolding are candidates. Be specific about the signal, the pivot, and what you learned. |
| behavioral_situational | Give me an example of using quantitative data to override strong engineering intuition on a platform product decision. | This is behavioral question #8 verbatim. Felix explicitly mentions using SQL/BigQuery telemetry to prioritize developer pain points at Intuit. The interviewer wants a specific story with numbers, not a general claim about being data-driven. | Use your ICE telemetry work: what the engineering team believed about developer behavior, what the usage data actually showed, and how you used that data to change the roadmap. Be specific about the metric, the magnitude of the discrepancy, and the outcome of the data-driven decision. |
| behavioral_situational | Tell me about a time you had to deliver quickly while maintaining a high bar for quality on a developer-facing product. What did you cut, what did you protect, and how did you communicate those tradeoffs to stakeholders? | The role explicitly calls for 'deliver quickly while maintaining a high bar for product quality.' Felix's Splunk Scheduler delivery (4 months end-to-end) and Mailchimp GCP-to-AWS migration are strong candidates. The interviewer wants to see his quality bar, not just his speed. | Use the Splunk Scheduler story: what the original scope was, what you cut to hit the .conf19 demo deadline, what quality gates you refused to compromise (API contract, error handling), and how you communicated the scope reduction to stakeholders. |
| role_specific_scenario | OpenAI is shipping GPT-5.3, 5.4, and 5.5 in rapid succession. A new model capability — say, native tool-call parallelism — changes the optimal design of the agent execution API you shipped 3 months ago. How do you think about versioning, deprecation, and developer communication in that scenario? | Interview question #9 about backward compatibility is directly triggered by OpenAI's actual model cadence. Felix's ICE rSocket migration experience is relevant, but the interviewer wants to see how he'd handle this at OpenAI's specific pace and developer ecosystem scale. | Prepare a concrete versioning framework: API versioning strategy (date-based vs. semantic), deprecation timeline policy, migration guide requirements, and how you'd use the developer ecosystem (Campus Network, forums) to communicate. Reference your Intuit experience with SDK migration as a concrete anchor. |
| role_specific_scenario | Where should the boundary be between what an AI agent can do autonomously versus requiring a human-in-the-loop checkpoint — and how would you encode that into product decisions for OpenAI's Agents API? | This is culture/values question #10 verbatim and is directly relevant to OpenAI's safety mission. Felix's AutoEval system (PASS/FAIL reports with confidence scores) and his Fintellect financial advisory agents both involve consequential autonomous actions. The interviewer will probe whether his safety thinking is principled or superficial. | Develop a principled framework: irreversibility of action, scope of external side effects, confidence threshold, and user context. Use your Fintellect financial agents as a concrete example — what actions you required human confirmation for and why. Connect to OpenAI's stated safety mission without being generic about it. |
| motivation_fit | You have a NeurIPS publication, you've hand-coded BPTT in C++, you've built RL post-training workbenches, and you're running two AI startups. Why a PM role at OpenAI rather than a research or engineering role — and why now? | Felix's profile is unusually technical for a PM candidate. The interviewer will probe whether he's settling for PM or genuinely motivated by the product leadership dimension. The 'why now' question also surfaces the startup-to-BigCo transition motivation. | Be direct: articulate what you can do as a PM at OpenAI that you cannot do as a researcher or engineer — specifically, the ability to shape what gets built and for whom at the platform layer, with the leverage of OpenAI's developer ecosystem. Avoid 'I want both worlds' — pick a lane. |
| motivation_fit | OpenAI's mission is ensuring AGI benefits all of humanity. Your Fintellect platform targets retail investors and your Streamio platform targets real estate and financial markets. How does your work connect to that mission, and where do you think the API Agents platform specifically advances it? | OpenAI's culture/values questions probe mission alignment, not just product fit. Felix's fintech and real estate work is commercially motivated — the interviewer will want to see whether he has a genuine perspective on the broader mission, not just a rehearsed answer. | Connect your Fintellect work (democratizing financial literacy for retail investors) to the mission authentically — this is a real example of AI benefiting underserved users. Then articulate a specific view on how the API Agents platform advances the mission: by enabling the next generation of developers to build transformative applications, not just enterprise incumbents. |
| unique_to_this_interviewer | Note: No individual interviewer profile was provided — the INTERVIEWER input was the question bank itself. This question is therefore anchored in the question bank's implied perspective. The question bank shows a strong bias toward API primitive design and SDK ergonomics. Given that, here is the highest-signal question not yet covered: You've built streaming pipelines (HLS, SSE, WebSocket) and you've built SDK interfaces (Java/Python Starter Kits at Intuit). If you were designing the streaming agent run events SDK method for OpenAI's Agents API today, what would you do differently from the Responses API's current streaming design, and why? | The question bank's SDK question (#3) is the most technically specific and differentiating question in the set. Felix has direct implementation experience with both streaming infrastructure and SDK design. An interviewer with this question bank is probing for PM candidates who can engage at the interface design level, not just the product strategy level. | Study OpenAI's current Responses API streaming design (event types, SSE format, error handling). Then prepare a specific critique and improvement proposal grounded in your own streaming implementation experience — focus on backpressure handling, event ordering guarantees, and error taxonomy gaps you've encountered in practice. |
Preparation priorities
- 1. SYSTEM DESIGN DEPTH: Prepare a complete, whiteboard-ready design for a durable, resumable agent execution API — primitives, state management, partial failure handling, and idempotency. Anchor it in your OpenClaw architecture but extend it to an external public API surface. This is the highest-signal question in the set.
- 2. BEHAVIORAL STORIES WITH SPECIFICS: Prepare 3-4 tight STAR stories from Intuit and Splunk covering: (a) driving alignment on a technically ambiguous decision, (b) a wrong abstraction you shipped and corrected, (c) data overriding engineering intuition. Each story needs specific metrics and honest acknowledgment of what went wrong.
- 3. FOUNDER-TO-PLATFORM-PM TRANSITION NARRATIVE: Develop a crisp, honest answer to why you're returning to a large org, what you'll find harder, and why OpenAI specifically — not a generic 'I want scale' answer. The 9-month founder gap will be probed in every conversation.
- 4. OPENAI PRODUCT FLUENCY: Get deeply familiar with OpenAI's current Agents SDK, Responses API streaming design, and the Codex agentic platform. You need to have a specific, informed opinion on what's right and wrong about the current developer experience — not just general agentic infrastructure opinions.
- 5. SAFETY AND HUMAN-IN-THE-LOOP FRAMEWORK: Develop a principled, non-generic framework for autonomous vs. human-in-the-loop agent decisions. Ground it in your Fintellect financial agents and AutoEval work. Connect it explicitly to OpenAI's safety mission without sounding rehearsed.
⚠ Watch-outs
- WATCH OUT — FOUNDER DEPTH VS. CROSS-FUNCTIONAL PM DEPTH: Felix's recent work is largely solo/founder-mode building. If asked about driving alignment across research, engineering, and GTM, he must anchor in Intuit/Splunk stories — not Streamio or Fintellect. Attempting to use founder stories for cross-functional alignment questions will read as a gap, not a strength. Handle by explicitly bridging: 'In my founder work I've had to wear all those hats simultaneously, which gives me empathy for each stakeholder — but the most relevant example of cross-functional alignment at scale is...' then pivot to Intuit.
- WATCH OUT — WRONG ABSTRACTION QUESTION IS A TRAP: Question #7 (wrong abstraction you shipped) requires genuine intellectual honesty. Felix has shipped a lot of platform work, and the temptation is to give a sanitized, low-stakes example. An interviewer with this question bank will push back on anything that sounds like a humble-brag. Prepare a real example where the abstraction was genuinely wrong, the signal came from developer behavior data, and the correction required admitting the mistake to stakeholders. The Asterias GraphQL API or SDK Starter Kit scaffolding templates are candidates worth examining critically.
- WATCH OUT — TECHNICAL DEPTH PERCEIVED AS ENGINEERING, NOT PM: Felix's RL workbench and BRAIN rewrite are impressive but could trigger the question 'why aren't you an engineer or researcher?' He must consistently frame his technical work through a product lens — what developer problem it solves, how he'd prioritize it against competing needs, and what he'd cut if resources were constrained. Avoid letting technical discussions drift into pure engineering depth without anchoring back to product decisions.
- WATCH OUT — SAFETY QUESTION REQUIRES GENUINE PERSPECTIVE, NOT PLATITUDES: The human-in-the-loop question (#10) is a values probe at OpenAI, not a product design question. Generic answers about 'balancing autonomy and safety' will fall flat. Felix needs a specific, principled framework with real examples — his Fintellect financial advisory agents (where consequential financial decisions are at stake) and his AutoEval system (PASS/FAIL with confidence scores) are the strongest anchors. Connect explicitly to OpenAI's mission without sounding like he's reading the company's about page.