FROM: A10X FIELD DESKTO: FORWARD DEPLOYED ENGINEERSRE: BRIEF №2 · THREE MOVESDATE: JULY 2026 · WK 31
Brief №2. Three moves as always: what changed, why it matters to an FDE, and what to do this week. No link dumps. This week the field published its receipts, and they line up unusually well.
HiringAWS put a billion dollars behind the FDE role, and postings are up over 1,000%
What changed. AWS announced a $1B investment to embed engineers directly inside enterprise customers, joining OpenAI and Anthropic in the same race. Postings for forward deployed engineer roles are up more than 1,000% year over year. Senior FDE total comp at the frontier labs is running $560,000 to $785,000, while Palantir's median forward deployed engineer sits near $215,000. Cursor's global forward deployment lead is publicly calling it the hottest job of 2026.
Why it matters to an FDE. Look at that pay spread. The gap between $215K and $785K is not years of experience, it is proof. The labs pay the top of the band for people who can walk into a customer's mess and come out with something running in production, because they have more demand than they have engineers who can do it. Demand is not the constraint in this market. Evidence is.
→ DO THIS WEEK: pull three live FDE postings (one frontier lab, one infra vendor, one vertical AI startup) and highlight every line that describes an outcome rather than a technology. Those lines are the interview. Then write one paragraph per posting on the last thing you shipped that maps to it. If you cannot fill all three, you just found your next project.EvidenceJohns Hopkins measured frontier agents on real enterprise work: 28% first pass, under 8% repeatable
What changed. Before deploying anything, Johns Hopkins Medicine benchmarked frontier agents in a simulated environment with 25 applications, 77 tools, and tasks grounded in thousands of pages of managed care policy. Leading frontier agents completed about 28% of complex tasks on the first attempt. Fewer than 8% stayed successful across repeated runs. The dominant failure mode was not hallucination and not integration. It was reasoning through policy-dense situations. And the same model, moved between two harnesses, went from workable to failing outright on an end-to-end prior authorization task.
Why it matters to an FDE. This is the clearest public measurement yet of the thing an FDE actually gets paid for. In their words: the model is necessary but not sufficient, and the harness built for the workflow is what turns a capable model into a deployable agent. Model choice is a procurement decision any VP can make in an afternoon. Harness design is the job. Note the second number too: 28% once versus 8% repeatably. Anything that worked in a demo is close to a coin flip. Consistency across repeated runs is the number that decides whether a thing ships.
→ DO THIS WEEK: take any agent you have built and run the same 10 tasks five times each. Report first-pass completion and repeat-consistency as two separate numbers. If you have never seen that second number for your own work, you have been grading yourself on the demo, which is exactly what your customer will stop doing in week two.PlaybookAWS published the eval-gated deployment pipeline, with real before-and-after numbers
What changed. AWS shipped an end-to-end blueprint built with Motorway: five stages, each a gate that blocks the release. Build-time evals (tool selection above 95%, LLM-as-judge helpfulness above 85%), staging validation on synthetic traffic, shadow mode against real production traffic for at least four hours with a 2% deviation threshold that auto-pauses the rollout, A/B at 5% of live traffic, then full rollout with continuous online evaluation. The results: tool selection went from 1 wrong answer in 8 to 1 in 50, and time to detect an issue went from about four hours to minutes. Their suite grew from 50 test cases to 150 in three months, because every production incident became a new case.
Why it matters to an FDE. Brief №1 called shadow mode the pattern that survives the demo-to-prod cliff. It now has a reference architecture, published thresholds, and a named customer attached to the numbers. This is the artifact you bring to day one of a deployment. You are not promising to be careful, you are handing over five named gates, each with a number that blocks a release, and a deviation threshold that pauses the rollout without waiting for a human. Steal their sharpest detail as well: include negative cases that assert which tools the agent must not call. One-sided evals create one-sided optimization.
→ DO THIS WEEK: write the gate table for something you have already built. Five rows. Each row gets a metric and the exact number that blocks the release. One page, no code. That page is a stronger portfolio piece than another demo video, and it is the document that gets you treated as the adult in the room.The through-line
The receipts arrived this week, from both directions. A billion dollars of embedded-engineer hiring on one side, and a benchmark on the other showing frontier agents finish 28% of real enterprise work on the first try and under 8% of it reliably. Those are the same fact stated twice. Companies are buying harnesses, not models, and they are paying whatever it takes to find the people who can build them.
See you next Monday.
— THE A10X FIELD DESK
Sources: HR Executive, AWS joins OpenAI and Anthropic in race to embed engineers inside enterprises; Healthcare IT News, Why Johns Hopkins is benchmarking AI agents before deployment; AWS, Evaluating AI Agents: a production blueprint with Strands and AgentCore.