What is in the customer layer an FDE can change without review?

Prompts and the customer's vocabulary, business rules, thresholds, routing, connector settings, and the customer's eval rows. The FDE changes it the same day, because a prompt change that needs a ticket and a release train means the company has a queue with a travel budget rather than FDEs. Two conditions: the layer is versioned, and it diffs against the shipped default in one command, so the delta in the handback packet is generated rather than remembered.

Why does a harness change need a pull request and a price?

Because the harness sets what every answer costs. HarnessTax ran seven models through Claude Code, Codex CLI, and Pi on SWE-bench Lite and Terminal-Bench 2.0. Claude Fable 5 solved 97.8 percent of attempts in Claude Code and 96.7 percent in both Codex and Pi, while Claude Code cost about twice Pi per attempt, $1.33 against $0.67. Pi reached the Pareto frontier with four tools: read, write, edit, and bash. Every tool description an FDE adds is tokens on every call for that customer. The PR carries the customer's eval rows that motivated the change and the cost per case before and after, measured in the sandbox, and engineering decides within two weeks: generalize, keep as a per-customer skill, or decline.

Can an FDE fork the core product for one customer?

No. A fork is a confession that the customer layer is too small, and the fix is a layer-one or layer-two pull request that moves the boundary for everyone. The one exception is an FDE embedded with a product squad for a week to move that boundary, which is product engineering with a visitor badge rather than deployment. The handbook test still applies: if any of the last three things an FDE built needed a core PR, the boundary is in the wrong place.

What does engineering owe the FDE in return?

A per-customer configuration that is versioned and diffs against the shipped default in one command. A sandbox close enough to the customer that a cost per case measured there means something, the sandbox rather than the mud in Neej Gore's phrase. An eval run on every customer's rows on every harness change, because per HarnessTax an alternative harness had the highest success rate in nine of twelve model comparisons, so harness changes can move quality as well as cost. Cost per resolved case, per customer, on the dashboard next to resolution rate. And a review window of two weeks at most, owned by a named engineer.

How did Palantir draw the line between FDEs and product engineers?

Per Nabeel Qureshi's Reflections on Palantir, forward deployed engineers solved the customer's problem without worrying about overfitting, and product development engineers took what they built and generalized it. FDE code got the job done fast, which he politely calls technical debt and hacky workarounds; PD code scaled cleanly. The line was easy to keep because the two groups wrote different artifacts. In an agent product the line runs through one prompt, one skill, or one tool description, which is why it has to be written down as a contract.

Is Ramp's FDE team inside engineering?

Yes. Leo Mehr's account has the team growing from two engineers to sixteen in eighteen months, inside engineering, working on the core product for enterprise customers, with one rule worth borrowing for any FDE pull request: always be scoping. Who is driving the urgency, who actually uses it, and will other prospects need it. He also describes the team heading toward spending most of its time operating harnesses and maintaining evals, which makes the PR-with-a-price rule the job rather than a constraint on it.

What a Forward Deployed Engineer Can Change in the Codebase

What can a forward deployed engineer change in the codebase?

THE SHORT ANSWER

Three layers, three rules. The customer layer (prompts, business rules, thresholds, routing, connector settings, and the customer's eval rows) belongs to the FDE: changed the same day, no review, on two conditions, that it is versioned and that it diffs against the shipped default in one command. The harness (tool set, agent loop, context assembly, guardrails, model choice) belongs to engineering: an FDE changes it only by pull request, and the PR carries the customer's eval rows and the cost per case measured in the sandbox, with a decision inside two weeks. The core belongs to engineering only, and the FDE never forks it. The reason the harness needs a price is HarnessTax, the UC Berkeley study published September 16, 2026 by Melissa Z. Pan, Shuo Yang, Negar Arabzadeh, Wei-Lin Chiang, Ion Stoica, and Matei Zaharia: the same model reached near-identical success rates across Claude Code, Codex CLI, and Pi while costing up to 5x more depending on the harness, and Tomasz Tunguz counted no statistically significant quality difference in 42 comparisons. Every tool and skill an FDE adds on site is a harness change, so it is a margin decision for that customer. Engineering owes back a configuration that diffs, a sandbox where a cost number means something, an eval run on every customer's rows on every harness change, and cost per resolved case per customer on the dashboard.

The line between what a forward deployed engineer may change on site and what belongs to engineering used to be easy to see, because the two groups wrote different kinds of code. In an agent product the line runs through a prompt, a skill file, or a tool description, and a UC Berkeley study this month showed that layer decides what every answer costs. That turns each on-site change into a margin decision, which is why the contract has to be written down.

Part 4 of The FDE Transition writes it down in three layers: what the FDE owns outright, what changes by pull request with a price, and what never forks, plus the four things engineering owes back. The handbook chapter on what has to change in the product before FDEs can work is where the configuration boundary comes from, and the deployment-to-product loop is what the pull requests feed.

Last reviewed 2026-09-27 · 1 min read

What can a forward deployed engineer change in the codebase?

SOURCES

THE LONG VERSION

RELATED ANSWERS