Welcome to the first A10X Brief. Every issue is three moves: what changed, why it matters to an FDE, and what to do this week. No link-dumps. A working engineer's read on where the edge moved.
NewThe gap that's stalling agent projects is evals, not frameworks
What changed. The 2026 State of Agent Engineering data is in: 57.3% of organizations now run agents in production, up from 51% a year ago. But the capability gap moved. 89% of teams have observability wired up — only 52% have evals. And the #1 barrier to shipping, cited by 32% of teams, is quality. Framework choice has basically commoditized (LangGraph leads production deployments — Klarna, Cisco, Vizient — but "which framework" is no longer the hard question).
Why it matters to an FDE. This is the single clearest hiring and value signal in the field right now. Companies have agents running and they can see them (observability), but they can't prove they work(evals). That gap is exactly what a Forward Deployed Engineer is hired to close. "I can stand up an agent" is table stakes. "I can prove it works before your customer finds the failure" is the $400K sentence.
→ DO THIS WEEK: Take one agent you've built and write three evals for it: (1) a tool-calling eval — does it call the right tool, with the right inputs, in the right number of steps; (2) a task-completion eval on 10 representative inputs; (3) one regression case from a real failure. That's a portfolio artifact and an interview answer in ~2 hours.ShiftEvals themselves shifted: single test cases are out, traces are in
What changed.The definition of "an eval" moved in 2026. Single test-case checks are no longer the source of truth. Autonomous agents now need trace-based evaluation (judging the whole trajectory of reasoning + tool calls, not just the final answer), production sampling (grading real traffic, not a frozen test set), and regression datasets that grow as new failure modes appear. Tool-calling evals specifically matter more than ever because agents now routinely connect to dozens of tools through MCP servers — more surface, more ways to call the wrong thing.
Why it matters to an FDE.If you show up to a deployment with a static 20-row test spreadsheet, you'll look a generation behind. The customers worth working with already have observability; what they lack is a harness that turns their production traffic into a growing eval set. Bringing that architecture is how an FDE becomes indispensable instead of a contractor.
→ DO THIS WEEK: Add a "reasoning" check to one eval: don't just grade the final output — grade whether the intermediate steps were relevant to the goal. Reasoning models sound deliberate even when the logic is wrong; catching that is a 2026 skill most engineers don't have yet.Field noteThe demo-to-prod cliff has a name now
What changed.The most common reason agent projects die is no longer a mystery — it's underestimating everything around the model: guardrails, evaluation harnesses, and a deployment story. The pattern the strongest teams use to survive the cliff: shadow mode before cutover. Run the agent against real traffic without acting, diff its outputs against the human baseline for a week, then cut over.
Why it matters to an FDE."It demoed great and died in prod" is the exact failure an FDE is parachuted in to prevent. Walking into week one with a shadow-mode rollout plan positions you as the adult in the room — before you've written a line of their code.
→ DO THIS WEEK: Draft a one-page shadow-mode rollout plan for your current project: what you'd log, what baseline you'd diff against, and the objective bar that triggers cutover. Keep it. It's the most reusable artifact an FDE owns.The through-line
The field didn't get harder to build in — it got harder to provein. Observability told everyone their agents were misbehaving; 2026 is the year the job becomes proving they don't. Evals, traces, shadow rollouts. That's where the leverage — and the salaries — are moving.
See you next Monday.
— THE A10X FIELD DESK