First, the disambiguation, because “Instinct” is one of the most overloaded product names in tech. AMD sells Instinct MI-series data-center GPUs. There’s a pet-food brand. There was a Pokémon Go team. This guide is about none of those. It’s about the AI assistant built by Spear Street Technology in San Francisco, founded in 2025 by a former Sierra research scientist named Noah Shinn. The product lets you reach an agent by phone or text, then hands off real tasks: drafting email replies, keeping a calendar straight, booking travel, arranging home services, and actually executing transactions on your behalf through WhatsApp, iMessage, and your inbox.
That last part is the whole interview, really. An assistant that answers questions is a chatbot. An assistant that spends your money and sends messages as you is a reliability problem with a very low tolerance for error. Everything Instinct screens for flows from that.
Who’s building it, and why that shapes the questions
Shinn is 23 and was first author on Reflexion (NeurIPS 2023), the paper about agents that summarize their own failures into memory and do better on the next attempt. He also worked on τ-bench, a benchmark for tool-using agents in realistic customer-service settings, and he was an early research scientist at Sierra, Bret Taylor’s enterprise agent company. If you know that background, you can predict the taste of the interview. This is a founder who thinks in evals, failure taxonomies, and tool-call correctness, not in leetcode-medium trivia.
The funding matters for a different reason. Instinct went from roughly $50M to a $2.5B valuation in about four months, with a $75M Series A led by Kleiner Perkins’ Mamoon Hamid, then a $250M Series B co-led by Index Ventures and Benchmark three weeks later. A company moving that fast, still in private beta, hires in tight loops and expects you to be useful in week one. Nobody there is running a six-week structured onboarding. Interviews reward people who can build a working agent flow end to end, not people who specialize in one narrow layer.
The likely interview loop
To be clear about what’s known: Instinct has not published an interview process, and at this size the loop changes month to month. What follows is inference, grounded in how comparable agent startups hire (a fellow consumer-facing agent company like Decagon screens hard for the same tool-use reliability), calibrated to the specific things this team is known to care about. Treat it as a prep map, not a leaked rubric. If your recruiter describes something different, believe your recruiter.
| Stage | Format | What it actually tests | Confidence |
|---|---|---|---|
| Recruiter / founder screen | 25-40 min call | Why personal agents, why now, why you; whether you’ve actually shipped an agent that touches real systems | High (universal at this stage) |
| Practical build round | 60-90 min, live or take-home | Wiring an LLM to tools, handling a bad tool response, retries, structured output that doesn’t break downstream | Medium-high |
| Systems / agent design | 45-60 min discussion | Designing a reliable action-taking loop: guardrails, confirmations, idempotency, latency budget over phone/SMS | Medium-high |
| Coding depth | 45-60 min | One real algorithm or a concurrency/latency problem, less LeetCode-puzzle than debug-this-live | Medium |
| Founder / team fit | 30-45 min | Ownership, speed, judgment about when an agent should refuse or ask a human | High |
The build round is the one that matters
If there’s a single round that decides the outcome, it’s the practical one. Expect a prompt shaped like a real Instinct feature. Something close to: “Here’s a tool that returns a user’s calendar and one that books a restaurant. Wire up an agent that takes ‘book me dinner Friday near the office’ and does it. Now the booking API returns a 500 on the second call. What happens?”
The naive version calls the model, parses JSON, calls the tool, done. The version that gets you hired handles the parts nobody demos: the model hallucinated a restaurant that doesn’t exist, the tool timed out, the user’s request was ambiguous about which office, the booking half-succeeded and you don’t know if you should retry. Talk about idempotency keys before they ask. Talk about when the agent should stop and text the user “did you mean the SF or NYC office?” instead of guessing. That instinct, knowing when confidence is too low to act, is the product.
A fair question if you haven’t shipped production agents: does a serious personal project clear the bar? It can, but only if it genuinely touches real systems and you can dissect how it broke. A weekend wrapper that calls one API on the happy path won’t. A side project that actually books, pays, or sends on your behalf, that you ran long enough to hit the ugly failures and can walk through them, reads a lot like real experience to this team. Realistically, though, the shorter the loop and the more senior the role, the more they lean toward people who have already carried an agent through live users.
Structured output comes up constantly here. Agents that call tools live or die on whether the model returns parseable arguments, so constrained decoding and schema validation are fair game. Have a real answer for what you do when the model returns almost-valid JSON at 2am in production.
Agent design, the way they’ll actually push on it
The design round is less “design Twitter” and more “design the thing we ship.” A strong prompt: sketch the loop that turns an incoming iMessage into a completed action, and defend where you put the guardrails. Good answers cover a few concrete tensions. Latency, because a user texting from their phone won’t wait 30 seconds for a reply, so you can’t run five sequential tool calls with a slow model for every turn. Safety, because the agent moves money and sends messages under the user’s name, so some actions need an explicit confirmation and some are safe to auto-execute, and knowing which is which is a judgment call. Recovery, because tools fail partway and the agent has to reason about what state the world is in now.
Evaluation is where Shinn’s background shows. Expect to be asked how you’d even know the agent is getting better. “It felt smarter” is a losing answer. Talk about task-completion rate on a fixed suite, regression sets built from real failures in the Reflexion spirit, and offline replays before shipping a prompt change. A concrete regression case is small: the exact user message, the tools that were available, the wrong action the agent took, and the correct action it should have taken, frozen so every future prompt change gets replayed against it. If you’ve read anything about how retrieval and agent systems get evaluated beyond a single accuracy number, bring it.
What they screen for, stated plainly
Underneath the rounds, a small fast-moving team is checking a short list. Can you ship a working agent flow without hand-holding. Do you have real judgment about reliability and when to defer to a human. Are you fast and low-ego, because the codebase will look different in a month and you can’t be precious about your abstractions. And do you genuinely care about the consumer-agent problem, because the people who join hot startups purely for the valuation tend to leave when the work gets grindy. Given the privacy and security scrutiny the product has already drawn, expect at least one pointed question about how you’d stop an agent from leaking or misusing personal data. Have a thoughtful answer, not a shrug.
Compensation, and how to actually check it
I’m not going to invent numbers for a private-beta company that hasn’t published bands, and you shouldn’t trust anyone who does. What’s true structurally: a Series B startup at a $2.5B valuation, backed by Benchmark and Index, competing with frontier labs for the same agent engineers, generally pays top-of-market cash plus meaningful early-stage equity. The equity is where the real variance lives.
To ground your expectations, pull current data rather than guessing. Cross-check base and equity ranges for comparable AI startups on Levels.fyi and Pave. When you get an offer, ask the questions that turn a headline number into a real one: is it options or RSUs, what’s the strike price against the latest 409A, what’s the preferred price investors paid, and what’s your rough ownership percentage after expected dilution. If you’re outside the US, two extra checks matter. Confirm visa sponsorship before you invest prep time, because at private-beta headcount many startups simply aren’t set up to sponsor yet, and it’s better to learn that on the first call. And treat early-stage private equity as harder to value and exercise from abroad, since the 409A strike plus the cash needed to exercise can turn into a real trap if you ever leave before a liquidity event. A $2.5B valuation makes options expensive to exercise and the paper value sensitive to the next round’s terms. Run a few scenarios through a total comp calculator, and read up on how to negotiate an offer at this stage before you counter. The cash is knowable; the equity is a bet, so price it like one.
How to prep in a week
Build a small agent that calls two or three real tools and handles one deliberately broken one, then write down every way it failed and what you’d do about each. That single exercise covers most of the build and design rounds. Read the Reflexion paper and skim τ-bench so you can talk about self-correction and tool-use evaluation in the interviewer’s own vocabulary. Keep your general coding sharp with a light study plan, but don’t over-index on it; this is not a company that filters primarily on hard algorithms. If you want the broader context, the AI-startup interview guides hub and the AI-era interviewing guide cover the patterns that show up across this whole cohort.
The thing to internalize is that Instinct isn’t hiring people who can describe an agent. It’s hiring people who’ve felt the specific pain of one calling the wrong tool at the wrong time and shipping anyway. If you’ve built that and have the scar tissue to show for it, most of this loop is just talking through work you’ve already done.
Practice the behavioral round:
