What FDEs Actually Do: AI Agents in Traditional Industries

Why I wrote this

On August 19, 2026, someone who identified themselves on X as an OpenAI Forward Deployed Engineer asked what technical topics people wanted their team to write about.

I asked: What does the role mainly do? What problems does it solve day to day? What skills are essential?

The reply was a Tech in Asia feature behind a subscription wall. These basic questions deserve a publicly readable answer, so I assembled one.

Question one: What does an FDE do?

My take: After listening to roughly 40 minutes of a Cresta presentation about FDE work, my understanding is simple: FDEs mainly bring AI agents into customer service, logistics, restaurants, and other traditional industries and existing workflows to make them more efficient.

It feels like a new wave of the internet changing traditional industries. The last wave moved information and processes online. This wave lets software retrieve information, use business systems, complete parts of a job within clear permissions, and hand work back to a person when needed.

One sentence: An FDE works with a customer to choose a workflow worth improving, connect AI or an agent to the systems already in use, put it into production, and prove whether it helps.

Question two: What problems does an FDE solve day to day?

Take customer support. An agent may need to verify the user, look up an order, read the refund policy, and update a system; when it reaches the edge of its authority, it must hand the full context to a person. An FDE's daily work is making that kind of workflow actually run:

• The customer says only, “We want AI.” The FDE has to turn that slogan into one specific, valuable, testable job—and sometimes say AI is unnecessary.
• The old systems do not fit together. Data may be messy, APIs incomplete, and permissions complicated. Much of the work is connecting them safely.
• The agent will sometimes be wrong. The team needs realistic tests and clear rules for which failures are acceptable and which require stopping or human review.
• Actions create consequences. Before an agent updates an order, issues a refund, or changes a record, it needs permissions, approvals, logs, and rollback.
• Building it does not mean people will use it. The FDE must train users, change the workflow, and prove that the system saves time, reduces errors, or improves an outcome.

A day might therefore include backend or full-stack coding, mapping a workflow with the customer, testing an agent, reading production logs, or fixing a launch issue. In public accounts, Baseten's Het Trivedi estimated that roughly 75% of his first six months went to engineering and model optimization. Cresta's Hanze Li said his coding time moved from more than half early on to less than one-third later. There is no fixed split; the work changes with the stage of the project.

Question three: What skills matter?

My technical comparison: Compared with my current backend work, I do not see greater pure technical depth in common FDE work.

The harder part is breadth: the same role often covers all of the following.

• Engineering: Write the code, connect the systems, and handle production failures.
• AI: Know when to use ordinary software, a simple model call, retrieval, or an agent—and test whether the model and tools are reliable.
• Business: Learn how customer support, logistics, restaurants, or another industry actually works.
• Communication: Ask frontline workers the right questions and explain cost, risk, and trade-offs to executives.
• Delivery: Cut low-value scope and ship a small version that genuinely helps.
• Outcomes: Define done as “people use it, results are measurable, and failures are manageable,” not “the code merged.”

For me, FDE looks more like broader ownership than a higher rung of pure technical depth. The rare part is combining engineering, AI, industry understanding, and customer delivery.

Conclusion

The last internet wave moved traditional workflows online. This wave puts AI agents inside those workflows so software can complete part of the job. FDE may not be deeper than backend engineering, but it sits closer to the industry, the customer, and the final result.

Sources: Role and agent details come from OpenAI's public job description, Presence, and Frontier. Day-to-day examples come from public employee accounts at Baseten and Cresta. My takeaway from the roughly 40-minute Cresta presentation and the technical comparison are my own views.