# Manus AI interview guide for engineers

Source: https://www.techinterview.org/post/3233477944/manus-interview-guide/
Updated: 2026-11-12 · techinterview.org

Manus publishes no interview rubric and has almost no candidate reports, so the loop below is reverse-engineered from its posted roles and Singapore base: expect a recruiter screen, one or two technical screens, and a final round, all weighted toward agent reliability (making a long chain of tool calls finish correctly) over LeetCode trivia. For context, Manus, the autonomous-agent product from parent company Butterfly Effect, raised more than $500 million in a round announced October 8, 2026, its first since Chinese regulators forced Meta to unwind its roughly $2 billion acquisition. No pay bands are published, so benchmark any offer against peer AI startups.

Manus sells an AI agent that runs inside its own cloud computer: give it a task and it plans, opens a browser, writes and runs code in a Linux sandbox, and hands back a finished deliverable instead of a chat reply. The hard fact about a product like that is reliability over long horizons. A single task can be dozens of tool calls deep, and one bad step poisons everything after it, so the people Manus hires have to reason about where an agent goes wrong and how you'd catch it.

The corporate story matters to a candidate because it changes who you'd actually be joining. Butterfly Effect, Manus's parent, [announced on October 8, 2026 that it had raised more than $500 million](https://techcrunch.com/2026/10/08/chinas-manus-raises-over-500m-in-first-funding-round-since-split-with-meta/), co-led by Boyu Capital and IDG Capital, with existing backers Tencent, HSG, and ZhenFund also in. The company said in a WeChat post that it "has raised more than $500 million," a reported raise, not a disclosed close with audited terms. The company has not disclosed a valuation; the $4 billion figure comes from [TechCrunch's September report](https://techcrunch.com/2026/09/18/manus-seeks-4b-valuation-in-new-500m-fundraise-as-it-resumes-independent-ops/) (relaying the Wall Street Journal) that Manus was in talks to raise at that number, not from the company.

## The Meta deal that was built and then unwound

This is stranger than a deal that simply fell through. Meta announced a roughly $2 billion acquisition of Butterfly Effect in December 2025. In April 2026, Chinese authorities ordered the deal unwound, and Meta was forced to call it off. Reporting is ambiguous about how much had completed before the reversal, and Manus later said it had to delete some user data as part of the split. February 2026 job postings still described Manus as part of Meta, so some integration had started before it was pulled apart, and the October round is the company standing back up on its own.

Many Manus postings on aggregators are from early 2026, filed under Meta's hiring org, so a req you read may predate the split and may no longer be open; confirm with the recruiter that the role is live and sits inside the independent company. Manus also relocated staff to Singapore in mid-2025 and is reportedly weighing a Hong Kong listing, so hiring, pay, and visa are Singapore questions, not Bay Area ones.

## What Manus builds, and why the loop screens for agent reliability

Each Manus session spins up an isolated Linux sandbox with a real browser (built on the open browser-use framework), a Python interpreter and file-system access, and a central orchestrator breaks a request into ordered steps and delegates to sub-agents such as a browser agent. There is also a Browser Operator extension that drives tasks inside your own browser tabs using your existing logins instead of a cloud browser, and Manus 2.0 shipped in the second half of 2026. These details come mostly from Manus's own docs and blog plus community reconstructions, so treat the internals as reported rather than confirmed line by line.

Manus mostly rents its intelligence: at launch in 2025 its chief scientist said it ran on Claude 3.5 Sonnet v1 and various fine-tuned Qwen models, and the Claude dependency reportedly tracked forward since, though Manus does not publicly pin one model. When the model is bought, the engineering that is left is everything around it. That is why the questions that fall out of the product are about evaluation (how do you score a 40-step task that has no single right answer), recovery (an action fails or the page changed, now what), context and memory (the agent keeps intermediate results, how do you keep that from blowing the token budget), and cost, since every step is tokens and sandbox time against a model you might swap out next quarter. A product like this lives in the same world as other agent companies such as [Harvey AI](/companies/harvey-ai-interview-guide/), [Decagon](/companies/decagon-interview-guide/), and [Cursor](/companies/cursor/), so prep your agent-systems fundamentals to that bar, not a generic web-app loop.

## The loop Manus hasn't published

There is no usable body of Manus candidate reports. It is young, the past year has been a whipsaw of an announced Meta takeover and a forced reversal, and its public footprint is mostly product and funding news. The table below is an inference for a company at this stage and in this market, not anything Manus confirmed, so read the stage count, order, and timeline as approximate and ask your recruiter for the real sequence before you prep. What matters more than the ordering is what each round tests.

For an agent or backend engineering role, expect a recruiter screen, then a technical screen that is practical coding in the job's language rather than exotic algorithms, then a deeper round that is part [system design](/system-design-interview-guides/) and part agent design: build and evaluate a tool-using loop, decide when it should stop, handle a step that fails. The posted AI agent engineer role lists building evaluation systems as part of the job, so expect that to show up as a design question rather than trivia. The loop likely closes with a final that bundles a harder technical round with a hiring-manager or founder conversation about how you actually work, where the usual [structured STAR answers](/post/3233460379/behavioral-interview-questions-2026-star-method-amazon-leadership-principles-and-winning-answers/) carry you. With the team in Singapore, some loops may run remote for overseas candidates, and the location on your req drives both pay and the work-pass question.

| Manus interview and company facts (as of November 2026) | Detail | How well-sourced |
| --- | --- | --- |
| What it does | Autonomous AI agent that runs tasks in an isolated Linux sandbox with a browser, Python interpreter, and file access; central orchestrator delegates to sub-agents; Browser Operator extension drives your own browser; Manus 2.0 shipped in 2026 | Reported (Manus docs/blog, community reconstructions) |
| Parent company | Butterfly Effect; staff relocated to Singapore in mid-2025; reportedly weighing a Hong Kong listing | Reported (press) |
| Meta history | Meta announced a ~$2B acquisition in December 2025; Chinese authorities ordered it unwound in April 2026 and Meta called it off; Manus deleted some user data in the split | Reported (TechCrunch, press) |
| Funding | More than $500M announced October 8, 2026, co-led by Boyu Capital and IDG Capital, with Tencent, HSG, ZhenFund and others; a $4B valuation was reported in September (WSJ, via TechCrunch) as a number Manus was in talks to raise at, never disclosed by the company | Reported raise (company WeChat post, TechCrunch); $4B sought, per WSJ/TechCrunch |
| Revenue | Annual recurring revenue reported above $100M around the Meta deal; later run-rate figures of $400M-$500M reported as a pace, not booked annual sales, and behind a paywall you cannot independently verify | Reported (TechCrunch; The Information, paywalled) |
| Model stack | At launch (2025) the chief scientist confirmed on X (surfaced partly through community reverse-engineering, relayed by the-decoder) that Manus used Claude 3.5 Sonnet v1 and various fine-tuned Qwen models; later versions reportedly track the Claude family forward; Manus does not publicly pin one model and is not a foundation-model builder | Reported (chief scientist's X post via the-decoder, plus community reverse-engineering); current stack unconfirmed |
| Open roles | AI agent engineer (the posting lists building evaluation systems), backend (one posting asked for Go), frontend, and security/infrastructure, mostly in Singapore; many postings date to early 2026 under Meta's org and may be stale | Posted roles, early 2026 (investor/aggregator job boards) |
| Interview loop | Recruiter screen, one to two technical screens, final round; roughly 2 to 4 weeks | Inferred from comparable startups, not company-published |
| Compensation | No pay bands published; benchmark against peer AI startups and the exact Singapore req | Not published |

The [AI-startup interview difficulty index](/ai-startup-interview-difficulty-index/) and the broader [AI-native company interview guides](/ai-startup-interview-guides/) give a fuller picture than Manus's thin public footprint.

## What each role family is likely asked

These are the questions I'd bet on, because they fall out of what the product has to do. Figure out which family your req leans toward and go deep there.

- Agent engineering: design a tool-using loop for a real task (book something, scrape and summarize a site, fill a multi-page form), make it robust to a step that fails or a page that changed, decide how the agent knows it is finished versus stuck in a loop, and say how you would score a run with no ground-truth label and catch regressions when the underlying model changes.

- Backend: one posting asked for [Go (Golang)](/post/3233474456/go-golang-interview-questions-2025-goroutines-channels-interfaces-error-handling-context-generics-concurrency-patterns/), and the real design question is the orchestration layer. You are scheduling thousands of concurrent agent sessions that each burn tokens and sandbox time, so how do you enforce a per-task token and compute budget without starving the legitimate task that genuinely needs 200 steps to finish?

- Frontend: stream a long-running agent's progress to the user, show intermediate steps and artifacts, and handle cancel, retry, and a task that runs for minutes without hammering the backend.

- Security and infrastructure: the Browser Operator extension acts inside the user's own browser with their existing logins, so the sharp question is how you let an autonomous agent act as a signed-in user without a prompt-injected web page turning those credentials against them. The cloud sandbox raises the more familiar isolation and resource-limit questions. If that is your track, the [security engineer interview guides](/security-engineer-interview-guides/) fit better than a product loop.

The depth here sits closer to the [AI-era interviewing](/ai-era-interview-guide/) bar than a generic algorithms round: you will argue about reliability, evaluation, and cost per task more than model internals.

## Pay, equity, and the Singapore question

Manus publishes no compensation bands. The company said it will keep hiring at home and abroad after the raise, but none of the postings I could verify carried a pay range, so benchmark instead: pull Singapore-specific data from NodeFlair and levels.fyi for the exact role and level, and run the result through a [total comp calculator](/total-comp-calculator/) before you compare offers.

Equity is where the real uncertainty sits. Your grant would be priced against a valuation the company has not disclosed, with a reported $4 billion figure that is not official, at a company whose parent was just forced to unwind a $2 billion sale and delete user data. That is a different risk profile from a clean, never-acquired startup, so pin down the option or RSU count, the strike or reference price, the vesting schedule, and any cliff, and size the grant to the risk rather than the headline. The usual [salary negotiation](/post/3233474669/salary-negotiation-2026/) moves apply once you have numbers. One wrinkle worth raising on the first recruiter call: ask whether your req sponsors an Employment Pass and what the timeline is. Singapore's [Ministry of Manpower](https://www.mom.gov.sg/passes-and-permits/employment-pass/eligibility) sets an Employment Pass minimum qualifying salary, currently S$5,600 a month for younger applicants and scaling up with age, rising to a S$6,000 floor for new applications from 2027, so a role has to clear that bar before a work pass is even on the table.