Interviews by company · Eightfold

Eightfold's AI-assisted round hands you a page-long system to build with your own assistant, and wants to see it run.

A senior engineer who went through Eightfold's AI-assisted coding round in April 2026 posted the problem statement in full. It is a system: API request logs arrive as files from many sources in different formats, a configuration maps their fields, the service computes usage metrics per client and per source, enforces rate limits read from a config file, and writes insights per source as JSON, while files are appended and sources come and go. The candidate used their own IDE and assistant, found the interviewers uninterested in anything but whether it ran, and passed. Eightfold describes the round itself on its engineering blog: it is called AI Design and Coding, it lasts 60 minutes, and any AI tool is allowed. A second senior candidate, in September 2026, got a stream-processing problem and a No Hire on that round.

What the round looks like

Eightfold's engineering blog post "Rethinking engineering interviews in the age of AI" (October 2025) describes the round, AI Design and Coding: 60 minutes, any AI tool of the candidate's choice, and a real-world problem of the kind that comes up in an early-stage product discussion, which the candidate designs, codes and demonstrates as a working prototype. There is no penalty for using AI. The post says the round is judged on outcome and reasoning, and names architecture, problem decomposition, clarity of thought, AI literacy and judgment in how AI is used.

The April 2026 report shows one instance in full. You open your own editor with your own assistant and receive a statement that reads like a design document. Logs are files in an input directory, one subfolder per source, in that source's own format; new files appear every few seconds, existing files are appended, sources vanish, new sources with new formats arrive. The requirements list a generic insights format, processing within a few seconds of a change, request counts and average latency per client, the top five endpoints per source, per-client limits in a config file with offenders flagged, new formats without code changes, and throughput in the millions of requests per hour. Two sample inputs follow, a JSON object with a list of requests and a CSV, with different key names for the same fields.

The candidate's account of the interviewers is unusually frank: a few years of experience, no questions about why, a focus on getting it to run. They passed and moved on.

The September 2026 report is shorter. In that senior loop the round was called "System Design + AI Coding" and came after a 90-minute low-level design round; the problem was to collect a stream of data, process it and act on its input. The candidate found the interviewer not very helpful, got a No Hire on the round, and did not say which tool they used.

What a statement this size is asking

Nobody builds all of that in an interview. The statement is an ordering problem: which parts are the core, which are layers, and which are claims you make rather than code you write. Our read of the text is three layers. Parsing and metrics: map both formats onto one record through the configuration and write the insights per source. Limits: windows, a limit per client, a flag in the output, the file reread so a change shows without new logs. Live updates: read only what is new, remember where each file was, forget what disappeared. Throughput and the five-second window are claims about the third layer.

With an assistant that can produce any one layer in minutes, the round comes down to whether you chose that order, checked each layer against the sample files before moving on, and can say what you left out.

What is confirmed, and what is not

  • Stated by Eightfold: the name (AI Design and Coding), the length (60 minutes), any AI tool, a working prototype to design, code and demonstrate, no penalty for using AI, and the criteria (outcome and reasoning, architecture, decomposition, AI literacy, judgment).
  • Reported once, in detail: the setup (own IDE, Claude in VS Code), the full statement with both sample inputs, the interviewers' focus on running code, the result.
  • Reported once, briefly: a September 2026 senior loop where the round followed a low-level design round and was a stream-processing problem.
  • Not known: whether levels below senior get the round (intern and new-grad candidates report a different coding screen, run by Eightfold's own AI interviewer), whether other candidates see the same problem, how much running code weighs against the stated criteria, and whether the frank description of the interviewers generalises.

How we would prepare

  • Practise cutting a page into an ordered list. Take a long statement, write five lines that name the layers and the order, and mark the one that forces a design decision. Do it before the first prompt.
  • Have a field-mapping pattern ready. A per-source configuration that names the format and maps raw keys onto internal fields is what makes new formats a config change; it is also the first thing to build.
  • Know what a fixed window means for a rate limit. Aligned to the clock, counted per client across sources, strictly more than the limit is over. Say the rule; the assistant will pick one otherwise.
  • Make incremental reading a design, not an afterthought.Byte offsets per file, a half-written last line left for the next pass, a header remembered for CSV. It is the part the five-second requirement is about.
  • Keep the introductions short. The candidate's own advice; the clock is the constraint.

Nyrion's Eightfold section starts with API Log Insights, built from this statement. JSON Lines and CSV mapped onto one record through a configuration, with insights written per source; then rate limits read from a limits file on every run, with fixed windows counted across sources; then incremental runs that read only the new bytes and forget vanished sources. The statement, the sample formats, the mapping configuration and the limits file are what the candidate posted; the exact shapes, the window rule and the incremental rules are ours.

How this round compares with other companies' versions, and what every version grades: the AI coding interview guide.

Frequently asked, briefly answered

What is the format?

Eightfold calls the round AI Design and Coding: 60 minutes with any AI tool you choose (its engineering blog names Cursor, ChatGPT, Claude and Copilot), in which you design, code and demonstrate a working prototype for a real-world problem. The candidate we read in detail, a senior engineer in April 2026, used their own local IDE with Claude in VS Code and had prepared an instructions file for it. They passed.

What was the problem?

A system, not a function: process API request logs from many sources written as files in different formats into an input directory, map their fields through a configuration, compute per-client request counts and average latency and the top endpoints per source, enforce per-client rate limits from a limits file, and write insights per source as JSON, all while files are appended, sources appear and disappear, and new formats arrive.

What do the interviewers look for?

Eightfold says the round judges outcome and reasoning: architecture, problem decomposition and clarity of thought, AI literacy and tool fluency, and judgment in how you use AI. The April candidate's account is blunter: the interviewers were a few years into their careers, did not cross-question the code quality or the thought process, and wanted it to run. Keep the introductions short, because the problem is big and the time goes.

Is the full statement available?

The candidate posted it in full, including two sample inputs: a JSON file with a timestamp and a list of requests keyed by api_key and latency_ms, and a CSV with client_id and response_time_ms, with a note that a configuration is needed to map the different key names onto internal fields.

How much of it can be built in the time?

Probably not all of it. The round is 60 minutes, and Eightfold asks for a working prototype you can demonstrate. Our advice, not something the candidate reported: decide an order, build the parsing and the per-source insights first, add the limits, and leave the live-update behaviour for last; say what you left out and why.

Practice problems tagged for Eightfold

Browse the catalog and open the Eightfold section. No login required to look; sign in when you want to solve one.