Deploying Agents at Scale: What Enterprise Adoption Actually Looks Like, Michael Rishi Forrester, AI Startup Advisor, Neo4j Startup Program. Agentic Engineering Days Zurich, 20 November 2026.
Twenty-five minutes held the first half: what the careful studies found when teams adopted AI coding tools, and why the evidence disagrees with itself. The talk sent you here for the second half, the concrete part. Each page gives its figures with a label saying what kind of evidence they are, the source, and the date it was read.
| Page | What it covers |
|---|---|
| What AI-generated code ships: the security data | What ships, at what rate and against what baseline, from benchmarks, user trials, studies of published code and vendors' data |
| The evidence index | Every figure in the talk and on these pages: controlled study, vendor telemetry, self-reported or forecast, with source, read date and status |
| Two platform teardowns | Two agent incidents of my own: a cluster an agent reset, and a workshop fleet an agent reported that did not exist |
| Review, testing and what to measure when agents write the code | Review as the bottleneck, what an agentic first pass changed, what the randomized trials found, tests filtered before a person sees them, and a measurement set in place of PRs merged, with who measured each part and what it showed |
| Guardrails in two layers | Reasoning-layer controls the agent reads, infrastructure-layer controls it cannot reach, and the evidence on which holds |
| Which skills the engineers needed | Review, testing, specification and systems thinking, and what breaks without each |
| Is your environment ready? A self-check, and what to fix first | Ten yes-or-no questions in fix-first order, each tied to the evidence behind it |
| The interactive self-check | The same ten questions in your browser, naming your first fix as you answer; it loads, sends and stores nothing |
Every figure on these pages was read in its primary source, and the date is given. Where a figure could not be read in a primary, the page says so and marks it unverified rather than leaving it out silently. No page describes any employer's or client's internal systems.