Bloomberg AI-Assisted Coding Round, All Levels: 14 Experiences

Shared by janebeingjane ·
Level
Unknown
Stages
Phone screen → Onsite

Since October 2026, Bloomberg interns, new grads and experienced hires have reported an AI-assisted coding round: fix bugs, close tickets or add features.

At a glance

  • Process: one or two AI rounds early in the loop, often a small file first and a large codebase second (Intern 02 and 05, New grad 02, Level unknown 02). Where the rest of the loop is reported, system design, HR or a manager round came next. One experienced engineer had two AI interviews inside an in-person onsite (Experienced 02).
  • Asked recently: fixing bugs in existing code (5 of 7 intern and 2 of 3 new grad reports), checking code against a vague spec, closing tickets in a large repo, and building features in a Django app until the tests pass.
  • Difficulty: ratings ran from easy (Intern 01) to hard (Intern 03, New grad 02), and reading unfamiliar code fast was the hard part for New grad 02 and Experienced 01. Two of four tickets was enough in one round (Intern 03), but finishing every task still ended in a no once (Level unknown 01).
  • Practise first: Debug a Stock Exchange Order Book, Resolve Tickets in Multi-Tenant Matchmaking, Debug a Ticket Booking System, Debug Seat Hold, Confirm and Cancel Rules and Build Three Features in a Watchlist and Alerts App.

Coverage: 14 software engineering candidates who took the round: 7 interns, 3 new grads, 2 experienced engineers and 2 whose level isn't given · places named include London, Warsaw, New York and Dublin · interviews and reports from October to November 2026, plus one undated reply

Each level has its own section below, with a short line per candidate and a link to the full write-ups. Entry numbers match the numbers in each level's own article.

What is Bloomberg's AI-assisted coding round?

It is a coding interview inside code you didn't write, with an AI assistant built into the editor. You don't start from a blank file, and there is no algorithm puzzle. You read existing code, find what is wrong or missing, and change it. The interviewer watches how you understand the code and how you work with the assistant. Candidates call it the AI interview, the AI-assisted coding round or the AI-assisted codebase interview.

When it started. The earliest firsthand report is an intern interview on Oct 6, 2026 (Intern 07). Two experienced candidates add timing in replies from early November 2026. One, whose loop ran from July to September 2026, had no AI round. The other went through a New York final onsite in fall 2026 and never had one either. Before the AI version, some Bloomberg loops already had code review and refactoring rounds without AI, reported from 2025 to July 2026.

Where the round sits in the loop

  • Interns: the first round (Intern 01, 02, 05, 06), sometimes followed by a second AI round in a large codebase (Intern 02, 05) and then system design (Intern 03, 05). Intern 03 had an on-campus first round, then a ticket round in a big repo.
  • New grads: both rounds of the technical phone stage (New grad 02), or the first phone screen (New grad 03). A London loop listed two AI interviews, system design, HR and an engineering manager, with the middle three on one superday, a single day of back-to-back interviews (New grad 01).
  • Experienced: the second round, after a first round the candidate didn't describe (Experienced 01), or two AI interviews inside an all-day, in-person onsite with system design and HR (Experienced 02).
  • Level unknown: a virtual round 1 and a virtual onsite round 2, on different days (Level unknown 02).

The setup: code, language, AI tool and time

  • Code size: one file of about 100 to 150 lines (Intern 04, 06; New grad 03), one file with its tests (Level unknown 02), or a large repository with many parts (Intern 03, 05; New grad 02; Experienced 01; Level unknown 02).
  • Language: Python in every report that names one (Intern 04; New grad 03; Experienced 01; Level unknown 02). Experienced 01's app used Django, a Python web framework. Experienced 02 called its tasks language-agnostic.
  • The AI tool: most reports just say an AI assistant. In Level unknown 02 the editor ran inside HackerRank and looked like VS Code with a Copilot-style agent. An agent is an assistant that edits files itself instead of only answering questions. That editor had three modes: plan (propose steps), ask (answer questions) and agent (make the changes). The candidate read each change the agent proposed and accepted or rejected it. Experienced 01 named Sonnet 5 and GPT 5.6 as the built-in agent's models. In Intern 06, too, the assistant could switch to an agent that changes files directly.
  • How much AI to use: this varied. Using the agent was required in Level unknown 02. Experienced 01 suspected afterwards that the interviewer wanted the agent to do everything. Intern 06 held back, worried that heavy use would count against them. Level unknown 01 never found out what was wanted. In an uncounted Nov 18 report (see Level unknown), the interviewer stopped the candidate from editing by hand and sent every change through the AI.
  • Time: about an hour where a length is given. Experienced 01's hour began with 30 minutes of behavioral questions. New grad 01's AI interviews were listed at one hour. Intern 03 closed tickets in about 30 and 10 minutes, and Intern 06 got a hint about 30 minutes in.

What came up

Most tasks were bug hunts in code someone else wrote. Counts are per level, because the levels are separate pools.

Task type Reports Last seen
Find and fix bugs in existing code Interns 5 of 7, new grads 2 of 3 Nov 2026 (report date)
Close tickets in a large repo Interns 1 of 7, new grads 1 of 3, level unknown 1 of 2 Nov 2026 (report date)
Check code against a vague or written spec Interns 2 of 7 Oct 2026
Build features until the tests pass Experienced 1 of 2 Nov 2026 (report date)

A ticket is a short work request, like a bug report or a small feature. Intern 03's ticket round never mentions an AI assistant. It is here because it matches the large-codebase second round that Intern 02 and 05 describe as AI-assisted. Four reports gave no task at all: Intern 02, New grad 01, Experienced 02 and Level unknown 01.

What interviewers judged

These are candidates' impressions. No report gives Bloomberg's scoring sheet.

  • How you use the AI: expected to use it for navigating, understanding, debugging and editing (Intern 04); for clearing up the spec and finding where code and spec disagree (Intern 05); explaining how it helped reach and check the fix (Intern 07); steering the agent and using its modes well (Level unknown 02). New grad 03 also felt AI use was judged.
  • How you explain: own test cases, the reasoning behind each change, data structures and running time (New grad 02); a clear talk-through (New grad 03); clarifying questions about the spec (Intern 04).
  • Whether the fixes hold: correct fixes and tests for edge cases (Intern 04). Intern 06 felt reaching a fix mattered more than the narration.

Do you need to finish every task?

Not always, but finishing isn't enough on its own either.

  • Partial work passed: Intern 03 passed with two of four tickets. New grad 03 passed and heard secondhand that someone else passed with four of about five bugs.
  • Partial work failed: New grad 02 closed two of four tickets and wasn't selected; the report doesn't say which round decided it. Intern 06 fixed one bug with the interviewer's help and didn't pass.
  • Full work failed: Level unknown 01 finished every task and was still rejected.
  • Outcomes by level: interns, one accepted offer (Intern 04), one still in process (Intern 02) and five without an offer; new grads, one pass of round 1 (New grad 03) and two without an offer; both experienced engineers rejected; level unknown, one rejected and one outcome not shared.

How sure is this? Small counts, and not every loop has it

The counts are small: 14 candidates in all, and only 2 experienced engineers. A pattern seen twice is a hint, not a rule. Several October and November 2026 loops had no AI round at all:

  • Interns: a London internship candidate (report posted Nov 16, 2026) had two one-hour LeetCode-style interviews on HackerRank, with no AI round mentioned.
  • New grads: a New York engineer listed at junior level, who interviewed in October 2026, discussed a project they had built, then got stuck on a string problem close to LeetCode 1209 (removing runs of repeated letters) and was rejected. No AI round is mentioned. It is the October report behind One-Dimensional Candy Crush in our new grad bank.
  • Experienced: an engineer with six years' experience (posted Nov 13) had a phone screen with a medium and a hard problem, a virtual onsite with two more, two design problems and HR, and got an offer. A senior New York candidate (posted Nov 18) was told AI was not allowed. The New York final onsite from the timing note above was all behavioral. Two more experienced candidates, posting Nov 16 and Nov 18, had classic coding phone screens; the second also had virtual coding and system design rounds. A senior London loop in September or October 2026 also had no AI round.

LeetCode-style rounds are still reported at every level. Our question banks list algorithm questions last seen in October 2026 for interns and new grads, and in November 2026 for experienced engineers. Reports with no level given have their own bank. For new grads, three AI-round reports and one without is a small sample.

How to prepare, from what candidates reported

  • Get a map before you edit. New grad 02's tip: ask the AI to describe the main objects, the components and how a request moves through them, then start changing code.
  • Treat the spec as a conversation. In Intern 04 the spec was vague on purpose and the interviewer played a project manager who answered questions.
  • Don't trust green tests. In Intern 06 bugs were still in the code even though both supplied tests were green. In Intern 01 the agent produced tests that fail on the bugs, and the fixes came after.
  • Practise steering an agent. Plan first, ask questions, and read every diff before you accept it (Level unknown 02). Experienced 01 worked one feature at a time in plan and ask modes.
  • Practise reading code fast while talking. Reading speed was the bottleneck in New grad 02's ticket round, and Experienced 01 found it hard to follow the agent's output while explaining it.
  • Know your data structures. New grad 02 had to explain running time and suggested a min-heap (a structure that always hands back the smallest item first) instead of sorting on every sell.
  • Prepare for the rest of the loop. System design was part of the loop in Intern 03 and 05, New grad 01 and Experienced 02, and LeetCode-style rounds still appear (see above). Intern 03's one piece of advice was plenty of mock interviews.

Practice on FastPrep

Each project gives you existing code and an AI assistant, like the real round.

  • Debug a Stock Exchange Order Book. Rebuilt from New grad 02's first round: the five bugs they reported, plus the speed follow-up as a discussion point. The report had no code, so the code is ours.
  • Resolve Tickets in Multi-Tenant Matchmaking. New grad 02's second round: the two tickets they described plus two practice ones. Also the shape of Intern 03's round.
  • Debug a Ticket Booking System. New grad 03's round, a short booking program with bugs. The report gave no code, so the bugs are ours.
  • Debug Seat Hold, Confirm and Cancel Rules. Built from Intern 04 and 06's reports: a booking class with three seat sections, hold, confirm and cancel rules, an open-ended product note, and two tests that pass while bugs remain.
  • Build Three Features in a Watchlist and Alerts App. Built on Experienced 01's round: features in an unfamiliar Django app until failing tests pass. One feature is a database change, as in the uncounted Nov 18 report. Neither report named an app, so the domain is ours.

Reported, not uploaded

These tasks were too thin to build: no code, domain or bugs were given.

Task Reported by Last seen
Locate bugs in an existing codebase through agent-generated failing tests, then repair them Intern 01 Nov 2026 (report date)
Find a spec's unclear points and fix the code that breaks it, small codebase then large Intern 05 Oct 2026
Root-cause an issue in an unfamiliar repo, fix and test it, explain the AI's part Intern 07 Oct 2026
Solve a task in one Python file that ships with tests, using the required agent Level unknown 02 —
Work through about five tickets in a larger Python codebase Level unknown 02 —

Interns: 7 reports

The round was the first or second step, and two loops had system design after it. All 7 intern reports in detail.

Intern 01: Failing tests from the agent point to the bugs

Report date: Nov 18, 2026 · Role: Software Engineer Intern · Applied: Online · Result: No offer

Round 1 gave them an AI agent and an existing codebase with bugs already in it. The bug hunt went test-first: the agent produced tests that fail where the code is wrong, and the code was repaired after that. They rated it easy. This may be a later report from Intern 02's candidate, since nothing recorded tells the two apart.

Intern 02: A small file now, a big codebase next

Report date: Nov 8, 2026 · Role: SWE Intern · Result: In process

Round 1 used one small file; the task wasn't shared. Round 2 was set to be AI-assisted in a large codebase, and the candidate worried about finding their way in it and about how the AI was meant to be used.

Intern 03: Two of four matchmaking tickets was enough

Report date: Nov 8, 2026 · Role: SWE Intern · Recruited: On campus · Result: Rejected after system design

After an on-campus first round came four tickets in a huge matchmaking codebase full of queues, workers and business rules. They closed two, in about 30 and 10 minutes, and passed; the report doesn't mention AI here. An open-ended system design round followed.

Intern 04: Seat holds, a vague spec and a project manager

Interview date: October 2026 · Role: Software Engineer Intern · Location: London · Result: Accepted an offer

The code was a Booking class covering three seat sections (VIP, Standard, Economy): about 100 lines of Python in one file. Holding should lower what is free, confirming must not take the seats twice, and cancelling gives them back. The spec was vague on purpose, and the interviewer played a project manager (practice).

Intern 05: Check the code against a spec, twice

Interview date: Oct 29, 2026 · Role: Software Engineer Intern · Location: London · Result: No offer

Both AI rounds paired a codebase with a written spec, small in round 1 and large in round 2. The spec was unclear in places and the code broke it in places. The interviewers wanted the AI used to move around the code, pin down the unclear points and spot where code and spec differ. Round 3 was system design.

Intern 06: Two green tests, bugs still hiding

Interview date: Oct 27, 2026 · Role: Intern · Result: Not successful

Bugs hid in a short seat allocation program (100 to 150 lines, plus a very short README) even though both of its tests were green (practice). Worried that leaning on the file-editing agent would count against them, the candidate used it sparingly. A hint came about 30 minutes in, and they fixed one bug together with the interviewer before time ran out.

Intern 07: An issue in a repo they had never seen

Interview date: Oct 6, 2026 · Role: SWE Intern · Location: Warsaw · Result: No offer

The repository was new to them, and something in it was broken. With the AI's help they traced the fault to its cause, then repaired and tested it. The interviewer also wanted a walkthrough of their thinking and of what the AI added, both in getting to the fix and in confirming it.

Not counted, possibly repeats. Three more intern reviews may describe candidates already listed, so they stay out of the counts.

  • Posted Nov 18: two AI rounds (the second in a much bigger codebase) and system design, plus two behavioral rounds, ending in an accepted offer. Nothing separates it from Intern 04.
  • Posted Nov 11, interview in November 2026, applied through the university: a short resume chat came with the AI task, and the code was a single source file paired with its tests, to repair first and then extend. Rated average; no offer. Nothing separates it from Intern 02 (whose round 1 was also one small file) or Intern 03 (also recruited through campus, no offer, round 1 content not shared).
  • Posted Nov 17: the first AI round felt easy, and nothing limited what the assistant would do, which the candidate found odd. No offer. Nothing separates it from Intern 02, 03 or 06.

New grads: 3 reports

New grad 02 and 03 met the round at the phone stage. The London loop (New grad 01) doesn't say exactly where its AI interviews fell. All 3 new grad reports in detail.

New grad 01: Two AI interviews and a London superday

Report date: Nov 16, 2026 · Role: SWE New Grad · Location: London · Result: No offer

The candidate listed five interviews: two AI-assisted ones, system design, HR and an engineering manager. The AI interviews were listed at one hour, design and the manager round at an hour each, and HR at 30 minutes. The middle three of the five fell on one day (a superday). The tasks weren't shared.

New grad 02: An order book, then four matchmaking tickets

Report date: Nov 13, 2026 · Recruited: Campus career fair · Result: Not selected

Round 1 was a StockExchange class whose order book maps each ticker to a list of sell orders (practice). Their bug list: float prices, order IDs that never advance, no input checks, a missing ticker check, and used-up orders left behind after partial fills. They also suggested a min-heap instead of sorting on every sell. Round 2 was four tickets in a game matchmaking repo; they closed two, a skill band that widens the longer a player waits and a lock that stops one player landing in two games.

New grad 03: A booking program with bugs, and a pass

Report date: Nov 8, 2026 · Recruited: Campus career fair · Result: Passed round 1

The first phone screen was a Python ticket booking program of a little over 100 lines with bugs planted in it, and AI help was allowed. They heard secondhand that there were about five bugs, and felt that communication and AI use were judged too.

Experienced engineers: 2 reports

Experienced 01 met the round in round 2 and Experienced 02 inside an in-person onsite; both ended in a no. Both experienced reports in detail.

Experienced 01: Django features in half an hour

Report date: Nov 9, 2026 · Background: Earlier jobs at two large tech companies · Roles: London and Frankfurt · Result: Rejected

Round 2 opened with 30 minutes of behavioral questions. Then came a large codebase built on Django, a framework they had never used, and a list of features to build until the tests passed (practice). They used the agent's plan and ask modes one feature at a time, and a few tests passed before time ran out.

Experienced 02: Two AI interviews in one onsite day

Interview date: Oct 22, 2026 · Level: Mid-level · Location: New York · Result: No offer

An all-day, in-person onsite held two AI coding interviews, system design and HR. The coding wasn't LeetCode; it was language-agnostic and tied to Bloomberg's own software. The tasks weren't described.

Level unknown: 2 reports

These candidates didn't say whether they were interns, new grads or experienced. Both reports in detail.

Level unknown 01: Every task done, still a no

Report date: Nov 9, 2026 · Role: Software Engineer · Location: Dublin · Result: Rejected

The candidate finished every task in the AI-assisted interview but couldn't tell how much AI use was wanted. They didn't describe the tasks or the stage.

Level unknown 02: A HackerRank editor with a required agent

Report date: Not given (a reply in a thread opened Oct 12, 2026) · Result: Not shared

Round 2, a virtual onsite on another day, was about five tickets in a larger Python project. Round 1, also virtual, was smaller: one Python file and its tests, with no framework and no spec. The agent was mandatory, and the candidate believed that steering it well, and picking the right mode, was graded.

Not counted, possibly a repeat. In a later round of a report posted Nov 18 (level not given), the job was one new feature in a full-stack Django app that needed a database schema change plus backend and frontend work. The candidate used Claude first to map the architecture and plan the work. When they then started editing code themselves, the interviewer stopped them: changes were to go through the AI. It may be Experienced 01, and Build Three Features in a Watchlist and Alerts App is the closest practice.

Common questions

Can you use AI in a Bloomberg coding interview?

In the AI-assisted round, yes: it is built into the editor, and in Level unknown 02 it was required. But some October and November 2026 loops had no AI round, and one senior candidate was told AI was not allowed (see "How sure is this?").

Which AI tool does Bloomberg's AI-assisted round use?

Most reports don't name it. In Level unknown 02 it was a Copilot-style agent in a VS Code-like editor inside HackerRank, and Experienced 01 named Sonnet 5 and GPT 5.6 as the agent's models.

Do you have to fix every bug in Bloomberg's AI round?

No. Intern 03 passed with two of four tickets. New grad 02 closed two of four and wasn't selected, and Level unknown 01 finished everything and was still rejected.

How long is Bloomberg's AI-assisted coding round?

About an hour where a length is given (New grad 01, Experienced 01). In Experienced 01, behavioral questions took the first 30 minutes of it.

Does every Bloomberg interview loop have an AI round now?

No. Intern, new grad and experienced candidates all reported October and November 2026 loops without one, and LeetCode-style rounds are still reported at every level.

Keep practicing

AI coding rounds at other companies

More companies with AI rounds on FastPrep:

Across companies: AI Coding OA Questions at 15 Companies: 32 Problems and AI Coding Interview Rounds at 43 Companies: 76 Problems.

Sources: Anywhere useful interview info can be found.

0 total reactions. Your reactions: none.

Comments

0

No comments yet — be the first to share what you know.