Company interview guide

Cognition Machine Learning Interview Guide (2026)

Hard take-home coding first, then practical rounds built around coding agents; very few first-hand reports.

Cognition's machine learning interview loop has 5 reported stages (Information session or recruiter contact, Coding assessment or take-home, Onsite: practical agent building and debugging, Agent evaluation, infrastructure and safety design, Behavioral round); typical length: Timeline not reported.

Roles covered: ML Engineer, AI Engineer, Agentic AI Engineer and Research Scientist. The question bank holds 9 questions reported from Cognition interviews.

Try a free question from this loop's topics

Interview process length
Timeline not reported
Stages
5
Reports linked to bank questions
11
Bank report dates
Sep 2026–Sep 2026
Questions in bank
9 (0 free)
Last reviewed
October 2026

How the Cognition interview loop works

In the five first-hand accounts (all Windsurf / Codeium ML engineer candidates, mid-2024 to early 2025) the process opens with a coding assessment: an hours-long online assessment or take-home with a window of three to four days, or a timed project of about three hours. One of those candidates also describes a group information session first, then an onsite of two technical rounds, a behavioral round and an offer. No first-hand account yet covers a Cognition (Devin) loop after the merger; a compiled public write-up describes practical agent building or debugging, infrastructure design and agent evaluation, plus a deployment scenario for Deployed Engineer roles. Research-engineer loops (post-training, mid-training, ML infrastructure) are not covered by any report we hold. The outline rests on few reports, so confirm every round with your recruiter.

The Cognition interview process, stage by stage

01Information session or recruiter contact

In one 2024 report, a group video session with several candidates; a recruiter contact is likely but not described

One Windsurf-era candidate was invited to a group information webinar with a handful of other candidates before the assessment. No account describes a one-to-one recruiter call, so use any early contact to ask how the current loop runs, since the company has merged two hiring processes.

What it tests

  • Basic fit and interest in the role
  • Which product and team you are aiming for (Devin, Windsurf, research)
  • Logistics: location, timeline, assessment window

How to prepare

  • Ask whether the first technical step is a timed online assessment, a multi-day take-home or a live round.
  • Ask which language and environment the assessment uses; one candidate disliked the assessment interface.
  • Have a short, specific answer on why coding agents and why this company.

A single 2024 Windsurf ML engineer report mentions a group information session before the assessment; no other account describes an opening call.

02Coding assessment or take-home

Either a hard algorithmic problem with a 3 to 4 day window, or a practical project of about 3 hours in a provided codebase with test cases

The best-supported stage. Two candidates got one hard algorithmic problem (dynamic programming over several constraints, or graph and tree traversal with cycles) where an optimal solution was expected and a slow one did not count. Two others got a practical task: extend a supplied codebase into a file replicator that copies directories between remote servers, passing the tests for each part, which one candidate said tested Python language features more than ML. A fifth called the coding assignment hard and time-consuming.

What it tests

  • Dynamic programming with several interacting constraints
  • Graph and tree traversal, including cycles and deep DFS logic
  • Reading and extending an unfamiliar Python codebase until tests pass
  • Careful file and directory handling across machines
  • Getting to an optimal solution, not just a correct one

How to prepare

  • Drill hard 2D DP and graph problems until you can reach the optimal complexity without trial runs.
  • Practise working inside an existing repository: read the tests first, then change only what each task needs.
  • Budget generously; candidates warn the assignment takes much longer than it looks.
  • Test edge cases yourself, since one candidate could not lean on running output when the solution was slow.

All 5 first-hand reports (Windsurf / Codeium ML engineer roles, 2024-06 to 2025-02, before the merger) open with this assessment: 2 hard algorithmic, 2 a file replicator project, 1 unspecified.

03Onsite: practical agent building and debugging

Two technical rounds in the one report that describes an onsite; the build-heavy content comes from a compiled write-up

One 2024 candidate reached an onsite of two technical rounds but did not describe them. A compiled public write-up of the current Cognition loop says it leans toward building: a hands-on round where you write or repair agent code. Expect questions on how an agent loop calls tools, recovers from failures and stays on task over long runs.

What it tests

  • Writing a minimal agent loop with tool dispatch and a step budget
  • Handling tool errors, timeouts and retries
  • Diagnosing an agent that drifts on long tasks
  • Designing tool schemas a model can use reliably
  • When multiple agents help and when they hurt

How to prepare

  • Build a small coding agent end to end before the interview so you can talk about real failure modes.
  • Read the company's public writing on multi-agent systems and be ready to argue both sides.
  • Keep a clear trace of every tool call in your code; it makes debugging live much easier.

One 2024 report confirms a two-round onsite without detail; the agent-building focus and the questions come only from a compiled public write-up, not first-hand accounts.

04Agent evaluation, infrastructure and safety design

Not confirmed as a separate round; may appear in the onsite. Deployed Engineer loops reportedly add a customer rollout scenario

No first-hand account describes a design round. The compiled write-up lists an infrastructure design round and an agent-evaluation round, with questions on running many cloud agents at once, why benchmark pass rates mislead, threat models for an agent with repository and CI access, and, for research roles, reward and environment design for agent RL.

What it tests

  • Evaluating a multi-step agent rather than a single response
  • Explaining gaps between benchmark gains and user experience
  • Threat modelling an agent with write access to code and credentials
  • Environment and reward design for long-horizon agent RL

How to prepare

  • Prepare an agent evaluation plan that goes beyond pass rates: task mix, partial credit, cost and human review.
  • Sketch sandboxing, secrets handling and network limits for an autonomous agent before the interview.
  • For research roles, be ready to discuss reward hacking and how you would detect it.

No first-hand report describes this round; it and its questions come from a compiled public write-up of the Cognition loop, so treat it as unconfirmed.

05Behavioral round

One round after the technical onsite in the single report that lists it

One Windsurf-era candidate had a behavioral round after the onsite and before the offer. Nothing is reported about its content, but the company's careers page stresses a small, talent-dense team working with urgency, so expect questions on ownership, pace and what you shipped.

What it tests

  • Ownership and pace on past projects
  • Motivation for coding agents
  • Working in a small, intense team

How to prepare

  • Pick two stories where you shipped something hard quickly and can explain the trade-offs.
  • Be honest about how you handle long, intense stretches; the company is open about its pace.

A single 2024 Windsurf ML engineer report lists a behavioral round before the offer; no other account mentions it.

Questions and themes reported most often

  • Practical project in a supplied codebase: build a file replicator that passes the tests (2 reports, coding assessment or take-home)
  • One hard algorithmic problem: dynamic programming over several constraints, or graphs and trees with cycles (2 reports, coding assessment or take-home)
  • Build a minimal agent loop with tool dispatch, error handling and a step budget (public lists) (onsite: practical agent building and debugging)
  • Reconcile the case against multi-agent systems with what works in practice (public lists) (onsite: practical agent building and debugging)
  • Evaluate an autonomous coding agent and explain why benchmark pass rates mislead (public lists) (agent evaluation, infrastructure and safety design)
  • Threat model for an agent with repository, CI and network access (public lists) (agent evaluation, infrastructure and safety design)
  • Environment and reward design for end-to-end agent RL (public lists) (agent evaluation, infrastructure and safety design)

What Cognition asks about

  • Take-home and algorithmic coding40%
  • Practical agent building and tool use25%
  • Agent evaluation, infrastructure and safety design20%
  • Behavioral and motivation10%
  • Agent RL and post-training5%

Cognition interview prep checklist

  • Solved several hard 2D DP and cyclic-graph problems to optimal complexity
  • Extended an existing Python codebase until all of its tests passed, within about three hours
  • Built a minimal agent loop with tool dispatch, retries and a step budget
  • Can explain how to evaluate a coding agent beyond benchmark pass rates
  • Sketched a threat model and sandbox for an agent with repository access
  • Asked the recruiter which rounds the current loop uses for your role

Tips

  • Treat the opening assessment as the main hurdle; every first-hand account starts with it and three of the five describe nothing beyond it.
  • Practise hard dynamic programming and graph problems to an optimal solution, not just a passing one.
  • Practise extending an unfamiliar Python repository against a test suite under a time limit.
  • Build a small coding agent yourself so the agent rounds feel like describing your own work.
  • Know Devin and Windsurf as products, and how an asynchronous cloud agent differs from an in-editor one.
  • Read the company's public posts on agent design before the onsite.
  • Ask your recruiter for the current loop; our reports predate the Windsurf merger and may not match it.

Practice the Cognition loop

Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.

Cognition interview FAQ

How many interview rounds does Cognition have for ML roles?

Candidates report 5 stages: Information session or recruiter contact, Coding assessment or take-home, Onsite: practical agent building and debugging, Agent evaluation, infrastructure and safety design and Behavioral round. In the five first-hand accounts (all Windsurf / Codeium ML engineer candidates, mid-2024 to early 2025) the process opens with a coding assessment: an hours-long online assessment or take-home with a window of three to four days, or a timed project of about three hours. One of those candidates also describes a group information session first, then an onsite of two technical rounds, a behavioral round and an offer. No first-hand account yet covers a Cognition (Devin) loop after the merger; a compiled public write-up describes practical agent building or debugging, infrastructure design and agent evaluation, plus a deployment scenario for Deployed Engineer roles. Research-engineer loops (post-training, mid-training, ML infrastructure) are not covered by any report we hold. The outline rests on few reports, so confirm every round with your recruiter.

How long does the Cognition ML interview process take?

Reported timelines run Timeline not reported.

What does Cognition ask in ML interviews?

By reported question mix: Take-home and algorithmic coding 40%; Practical agent building and tool use 25%; Agent evaluation, infrastructure and safety design 20%; Behavioral and motivation 10%; Agent RL and post-training 5%. The most repeated themes: Practical project in a supplied codebase: build a file replicator that passes the tests; One hard algorithmic problem: dynamic programming over several constraints, or graphs and trees with cycles; Build a minimal agent loop with tool dispatch, error handling and a step budget (public lists).

How should I prepare for the Cognition ML interview?

Treat the opening assessment as the main hurdle; every first-hand account starts with it and three of the five describe nothing beyond it. Practise hard dynamic programming and graph problems to an optimal solution, not just a passing one. Practise extending an unfamiliar Python repository against a test suite under a time limit.

Are there free Cognition interview questions?

LeetML has 9 questions reported from Cognition interviews. The stage sections above link the free practice problems for the topics each round covers; the reported-question set itself is in the app.

Related interview guides

Official sources