The Agent Engineer

The R&D engineering workflow has been structurally rewritten in 18 months. AI tools and agents can not only make engineers faster — they can change what engineers spend their time on, how long it takes to ship, what it costs per seat, and how reliable the output is. Wave 3 benchmarks indicate this functional transformation is scaling quickly, with many aspects now majority-adopted in production.

Automated coding and debugging in production jumped to 67% — up 30 points from Wave 2, one of the largest wave-over-wave jumps in our R&D benchmarks.

At 95% adoption, agentic-coding-tool access is reaching saturation. What Wave 3 reveals is a bifurcation in usage intensity that universal adoption numbers obscure: 53% of Runners have most or all engineers running parallel coding agents versus 8% of Walkers — a 6.6× gap. We believe this benchmark helps track the transition from individual engineering productivity gains to full R&D team workflow productivity gains.

The conventional 12-to-18-month B2B SaaS productization cycle appears to no longer apply. 71% of technical decision makers ship AI features from pilot to production in under six months; 25% do it in under three. Growth-stage organizations surveyed move fastest — 34% ship quarterly or faster versus 16% of enterprise peers. The engineering pipeline — evaluation, testing, deployment — has been compressed, likely due to the presence of AI automated tools often coding in parallel.

The per-seat software budget model is breaking in the age of AI. 33% of technical decision makers report per-engineer agentic-coding spend in the $251–$1,000 band, and 23% spend above $1,000 per engineer per month. Only 9% remain below $100. A 200-person engineering team at the modal band is spending an estimated $50K–$200K per month on agentic tooling alone — a P&L line item that did not exist 18 months ago. And 63% expect that spend to at least double in the next 90 days. In our view, it’s unlikely most engineering budgets have modeled this curve. We explore the P&L impacts of AI in Chapter 6: AI Meet P&L.

In Wave 2, agentic coding was clearly buying speed and leaving reliability behind. That gap has narrowed. Every engineering metric we track moved in the same direction this wave: the share of technical decision makers reporting a positive AI impact rose +17 points on development velocity and +15 on deployment frequency, but it rose fastest on the two reliability measures — +24 points on change-failure rate and +19 on mean-time-to-restore.

Reliability still trails in absolute terms. Development velocity is the standout at 87% positive, while change-failure rate (49%) and mean-time-to-restore (48%) remain the only metrics where fewer than half of teams report improvement. The direction of travel, though, is that the tooling is closing its own gap — and Chapter 7: The Governance Gap examines what is still missing from the controls around it.

11% of technical decision makers have no formal link between AI evaluation and optimization; another 4% perform no formal evaluation at all. Combined, 15% of engineering teams are deploying AI at production scale with no systematic way to detect degradation, capture failures, or improve model performance over time. At the current 67% production rate, that is not a small tail risk — it is, in our view, a meaningful share of the market operating without a feedback loop that every other production software system takes for granted.