.svg)
.svg)
2026 report for AI and engineering leaders
The new AI tech stack used by forward deployed engineers
Coding agents made code cheap. Discovery, LLM evaluation, guardrails and the customer's cloud are where AI projects now succeed or stall. This 2026 report maps the eight-layer FDE stack and the leading tools in each layer, with our estimate of how fast each one is being adopted.
Written by FullStack's AI Engineering Team. Published November 2026.
What you'll learn from this report
After reading this report, you’ll be able to:
See the whole AI agent stack. Map the eight layers FDEs work across, from discovery and AI coding agents down to identity and guardrails in the customer's cloud.
Know where AI projects really stall. Understand why LLM evaluation and AI observability now decide whether a pilot reaches production.
Sort the tools by layer. See the leading AI coding tools and agent frameworks layer by layer, from AI gateways to observability platforms, as of October 2026.
Benchmark your team. Compare your stack with FullStack's weekly-use estimates for each layer, built from several independent sources.
Prepare for an FDE engagement. Know what to have ready in each of the four phases, and the questions to ask any forward deployed engineer.
The shift
FDEs focus on ensuring outcomes - not coding
AI coding agents are now close to universal. JetBrains found 90% of professional developers use them every week, and we estimate about 95% of forward deployed engineers do. Writing the integration is no longer the slow part of an AI project.
The slow part has moved to the two ends of the AI tech stack. At the top, someone has to learn how the customer's work really happens. At the bottom, the system has to be tested against real data, then secured inside the customer's cloud and handed over. A study of more than 100,000 developers found lines of code rose about 741% with coding agents, while releases rose about 20%.
Forward deployed engineers work across all of it, inside someone else's security rules. This report maps the stack they use, layer by layer.
Explore the stack
The 8-layer AI tech stack, one layer at a time
Who this report is for
AI, engineering and IT leaders who are building FDE teams or hiring FDE partners, and want to know which tools and skills those engineers need to deliver results.
If you are weighing AI coding tools, LLM observability platforms or agent frameworks, the report shows where each one fits in the stack, and where no tool will help you.
From FullStack's CEO
Every company we talk to has an AI demo. Very few have an AI system that their operations team would miss if it went down on Monday. That second thing is what an FDE is for.
Table of contents
Five things to take away
The stack at a glance
Layer 1: Discovery
Layer 2: Agentic coding
Layers 3 and 4: Models, AI gateways and agent frameworks
Layer 5: Data and retrieval
Layer 6: Evals and observability
Layers 7 and 8: Deploying into the customer's cloud
The stack across the four phases of FDE work
Glossary, estimation method and sources
Two ways to get the report
Read it yourself. Enter your work email, and the PDF is yours. There’s no sales follow-up unless you ask for it.
Talk it through with us. Book a 30-minute call with FullStack. Bring the AI project in front of you, and we'll tell you which layers of the stack it needs, and which it can skip.
FAQs
What is an AI tech stack?
The set of tools used to build and run AI systems in production. For forward deployed engineers it spans eight layers: discovery, AI coding agents, models and AI gateways, agent frameworks and MCP, data and retrieval, LLM evaluation and observability, infrastructure as code, and the customer's cloud.
Which tools do forward deployed engineers use most?
AI coding agents such as Claude Code, Cursor, OpenAI Codex and GitHub Copilot are close to universal. Notebooks and work in the customer's cloud come next, followed by infrastructure as code. The report names the main players in all eight layers.
Do we need an AI gateway or an agent framework?
Not always. A small project on one model can call a cloud model endpoint directly. A gateway earns its place once you run several models, need spending limits or want one control point for security. The report explains where each fits.
Why do LLM evaluation and observability matter so much?
Language models can give different answers to the same question, so normal unit tests are not enough. Teams prove a system works with evaluation sets built from real cases, tracing and drift alerts. LangChain found only 37.3% of agent teams test on live traffic.
Who wrote the report, and where do the numbers come from?
FullStack's AI Engineering Team. Weekly-use figures are FullStack estimates. We built them from independent developer surveys and analyses of FDE job postings, among other sources. Every source is listed in the report.