AIAcademy · AIAcademy · 2026-05-16
2026 Agentic Coding Trends Report (PDF)
The headline number from Anthropic's 2026 Agentic Coding Trends Report is not the one most coverage picked up. It is not that engineers use AI in 60% of their work. It is the gap between that number and the share of work they actually delegate to an agent end-to-end: 0–20%, depending on task type.
That gap is the real literacy frontier of the next two years.
"AI-assisted" is a solved problem. Autocomplete, inline edits, pair-programming chat, a model reviewing a diff — all of it shipped in 2023–24 and is now ambient. The skills it requires are minor extensions of skills developers already had: read code, judge a suggestion, accept or reject. The cognitive cost is low because the human stays in the loop on every action.
"AI-autonomous" is a different job. You hand the agent an objective, walk away, and come back to a pull request, a deployed service, or a broken state you now have to debug. The skills required are not really programming skills — they are management skills: writing acceptance criteria, scoping work into delegatable units, placing checkpoints, knowing what to never delegate. The cognitive cost shifts upward, not downward: less typing, more specification and review.
Anthropic's report finds Cursor's own engineering team now has 30% of merged PRs opened by background agents. Nubank, Goldman, and Santander run "teams of Devins." The frontier ones — at the labs themselves — have crossed the delegation gap. Most of the industry has not.