The 2026 edition of the Theory AI survey reports a sharp shift in what blocks AI adoption. In 2025 the survey found that context limited AI performance; in 2026 that appears solved, and security is the dominant barrier. Based on replies from 455 senior technical builders, security and governance rose from sixth to first among pain points in twelve months. 43% of respondents now rate it the most important pain point, while evals are second at 22%. Current tooling value is still lower: only 10% say security tooling delivers the most value today, but 18% expect it to lead in the future, the largest expected gain of any area.

Model selection and cost optimisation have become standard practice. 45% of teams route queries across models, with 23.7% on a router they trained and 21.5% on a proprietary router; 54.7% still send every query to a single model.

The survey also shows concentration in agent sandbox infrastructure. Of the 259 teams writing code with agents, 62% host their sandboxes on Cloudflare, three times the closest competitor. The rest of the breakdown includes Other at 17%, Vercel at 17%, Modal at 16%, E2B at 12%, Daytona at 9%, and Sail at 7%. The write-up notes that the one control a builder cannot fake is concentrated in a single utility, and it is billed as compute rather than security.

For agent analysis, 56% of respondents use business intelligence tools to analyse agent performance. Coding agents also perform some of the analysis, at 48%, while dedicated observability tools are used by a roughly equivalent 41% and spreadsheets still survive at 30%. This suggests convergence in the observability category for this use case.

The full survey is said to contain more, including open versus closed source model adoption, the rise of MCP despite doubts last year, and the still early adoption of evals. The overall picture is one of rapid ecosystem evolution, with sophistication reaching the enterprise faster than any previous wave.