Grokking the Google Forward Deployed Engineer Interview
Vote

0% completed

​

What a Forward Deployed Engineer Does at Google Cloud

What the Role Is

Why the Proof of Concept Was Not Enough

What the FDE Built

How the Twelve Weeks Were Spent

Two Kinds of Output

How the Work Is Measured

Using This in an Interview

Key Takeaways

Practice Questions

Title image: one forward deployed engineer turns a 26-hour exception into a 5-hour exception, and every credit still waits for a person.
Title image: one forward deployed engineer turns a 26-hour exception into a 5-hour exception, and every credit still waits for a person.

A regional freight carrier runs 2,400 trucks and moves about 38,000 shipments a week. About 4,500 of them a week become exceptions: late, damaged, sent to the wrong depot, or billed wrongly. A desk of 60 agents resolves them. The median exception takes 26 hours, and the target is 6.

Three months earlier, a Google Cloud customer engineer had built a proof of concept, which is a small demonstration that shows an idea can work. Gemini drafted replies to 200 sample emails, and the desk liked them. The project then stopped. The demonstration could not read live shipment data, and the security team would not let an outside service into the network.

Google then assigned one engineer to the carrier for 12 weeks. At the end, the median exception took about 5 hours, and a desk agent still approved every credit to a shipper's invoice. That engineer was a forward deployed engineer.

What the Role Is

A forward deployed engineer (FDE) is a software engineer from a product company who works directly with one customer and builds what that customer needs for the product to deliver results in production. At Google Cloud the products are the Gemini models, Google's agent platform, and data services such as BigQuery. Google's postings call the FDE an "embedded builder" who closes the distance between frontier AI products and what works reliably inside a real company.

A fixed period of work with one customer, with an agreed goal at its end, is called an engagement. This engagement lasted 12 weeks. The FDE worked with the desk and with the carrier's platform team of 5 engineers, who would run the system afterward.

Why the Proof of Concept Was Not Enough

The proof of concept showed that Gemini could write a good reply, but it had no access to the carrier's data and no safe way to act on it. In the first two weeks, the FDE sat with 8 desk agents through 300 real exceptions and learned three facts.

  • The data lives in three places. Shipment status sits in the transport management system (TMS), an older database reachable only inside the carrier's network. Invoices and rates sit in a separate data warehouse. Exception emails arrive in a shared inbox.
  • The exceptions follow four types. About 52 percent are late, 23 percent are billing disputes, 15 percent are damage claims, and 10 percent are misrouted. Each type needs different data.
  • A wrong credit is expensive. A credit note lowers a shipper's invoice. A desk agent must approve every credit, and a supervisor must approve any credit over $1,000.

What the FDE Built

The FDE connected Gemini to the carrier's systems, split the work into narrow tasks, and kept every money-related change behind a person's approval. Google's GenAI FDE postings name each of these parts: multi-agent systems, MCP servers, and the "connective tissue" between Google's products and the customer's APIs, older data stores, and security perimeters.

The exceptions system: a coordinator agent per exception, three sub-agents, an MCP server inside the carrier's network, a read-only billing view, and a desk agent who approves every draft credit.
The exceptions system: a coordinator agent per exception, three sub-agents, an MCP server inside the carrier's network, a read-only billing view, and a desk agent who approves every draft credit.
  • An MCP server inside the carrier's network. The Model Context Protocol (MCP) is an open standard for connecting an AI application to outside tools and data. An MCP server is a small program that offers tools through that standard. This one has four read tools for the TMS and one tool that adds a note to a shipment. It runs inside the network, so the security team could approve it.
  • A read-only billing view. The agent reads only the invoice and rate columns it needs.
  • A coordinator agent with three sub-agents. An agent is a model that works in a loop: it calls a tool, reads the result, and decides the next step. A sub-agent is a second agent that receives one narrow task. The coordinator hands the shipment check, the billing check, and the reply draft to three sub-agents. The FDE built them with Google's Agent Development Kit (ADK), an open-source framework for agents.
  • A draft-only credit tool. The agent can create a draft credit note, but only a desk agent can issue it.
  • An evaluation set and tracing. An evaluation set is a fixed collection of past cases with known correct results. The FDE collected 600 past exceptions to score the system before it touched live work. Every exception also carries one id through every service, so a failure can be traced to the step that caused it.

How the Twelve Weeks Were Spent

WeeksStepWhat happened
1 and 2DiscoverySat with 8 desk agents through 300 exceptions; found the three data sources, the four types, and the credit rule
3 to 6BuildWrote the MCP server, the billing view, the coordinator, and the three sub-agents
7EvaluationOn 600 past exceptions, the agent's proposed resolution matched the desk's in 528, or 88 percent
8 to 10PilotHandled late and billing exceptions for 10 desk agents, with every credit approved
11 and 12Production and handoffHandled all four types for all 60 agents; the platform team took over operation

In evaluation, 24 of the 600 proposals, or 4 percent, contained a wrong credit amount. That number is why the approval step stayed in production.

Two Kinds of Output

A Google Cloud FDE delivers a working system to the customer and repeatable lessons to Google's product teams. The postings describe the FDE as a "critical feedback loop" who converts repeated field problems "into reusable modules or product feature requests." This engagement produced three items:

  • A reusable module: a template for MCP write tools that create drafts, not final records.
  • A feature request: a simpler way to connect a managed agent to a database inside a customer's network, which took two weeks to work around.
  • A deployment pattern: an evaluation set built from past tickets.
The two outputs of the engagement: a working exceptions system for the carrier and three items returned to Google's product teams.
The two outputs of the engagement: a working exceptions system for the carrier and three items returned to Google's product teams.

How the Work Is Measured

An FDE is measured by the customer's results in production, not by the number of agents or lines of code. For the carrier, three numbers mattered:

  • Median resolution time: from 26 hours to about 5, which is below the target of 6.
  • Desk time per exception: from about 22 minutes to about 9. At 4,500 exceptions a week, that is about 1,650 hours of work reduced to about 675.
  • Credits issued without approval: zero, by design.

Using This in an Interview

Interviewers want evidence that you can work inside a customer's systems safely and that you ship. Give the customer's numbers, and name the decision that kept the system safe.

"The carrier had a proof of concept that drafted good replies but could not read live data. I spent two weeks with the desk, then built an MCP server inside their network, a read-only billing view, and an agent with three sub-agents, with every credit held as a draft. On 600 past exceptions it matched the desk in 88 percent of cases, and in production the median exception fell from 26 hours to about 5."

Expect two follow-up questions:

  • "Why not let the agent issue credits directly?" In evaluation, 4 percent of its proposals had a wrong credit amount, and a wrong credit changes what a shipper pays.
  • "What did Google get from the engagement?" A reusable draft-only tool template, a feature request for private database connections, and an evaluation pattern that the next logistics customer can reuse.

Key Takeaways

  • A Google Cloud FDE works inside one customer's environment and builds what Gemini needs to deliver results there.
  • The postings name the work: multi-agent systems, MCP servers, connections to the customer's systems, evaluation, and tracing.
  • Discovery with the people who do the work decides which parts are needed and where a person must approve.
  • The FDE is measured by production results and also returns modules and feature requests to Google.

The next lesson, The FDE Teams and What Each One Ships, shows how this work differs across Google's FDE teams.

Practice Questions

1. The desk liked the replies from the proof of concept. Why did the project still stop?

<details><summary>Show answer</summary>

Because the demonstration could not read live shipment data, and the security team would not allow an outside service into the network. The FDE solved both problems by placing the MCP server inside the carrier's network.

</details>

2. Why does the agent create draft credits instead of issuing them?

<details><summary>Show answer</summary>

Because 4 percent of its proposals on past exceptions had a wrong credit amount. A credit changes what a shipper pays, so a desk agent approves each one before it is issued.

</details>

3. An FDE ships a system that cuts resolution time to 3 hours, but the carrier's platform team cannot operate it. Was the engagement a success?

<details><summary>Show answer</summary>

No, because the results stop when the FDE leaves. The handoff is part of the work, which is why the platform team took over operation during weeks 11 and 12, while the FDE was still there.

</details>

Reading Progress

0%


Vote for new content

On This Page

What the Role Is

Why the Proof of Concept Was Not Enough

What the FDE Built

How the Twelve Weeks Were Spent

Two Kinds of Output

How the Work Is Measured

Using This in an Interview

Key Takeaways

Practice Questions