60% Use, Only 0–20% Delegate — Anthropic's 2026 Agentic Coding Report Names the Real Bottleneck

Anthropic published its 2026 Agentic Coding Trends Report and the headline number is the one nobody on LinkedIn is bragging about. Surveying experienced engineers across companies that have shipped agentic coding into production, the report finds that developers use AI in roughly 60% of their work but say they can "fully delegate" only 0–20% of tasks. The gap between "I use it" and "I trust it to finish" is the real, dollars-and-cents bottleneck of 2026.

The report names eight trends, but three of them deserve to be printed and pinned to every engineering wall this week.

1. The SDLC Just Compressed from Weeks to Hours

Cycle time for a feature — spec, implementation, test, doc — collapses when an agent can do all four steps in one continuous pass. Teams that re-architect around this see a 5–10x cycle-time win. Teams that staple agents onto the old waterfall see a 5–10% win, and then ask why the hype was overblown.

2. Single Agents Are Evolving Into Hierarchical Multi-Agent Teams

Instead of one Claude Code session doing everything, the winning pattern is one orchestrator agent that fans out to specialized sub-agents — schema agent, query agent, security agent, doc agent — each with its own context window. This solves the "context rot" problem and makes agent runs auditable.

3. Long-Running Agents Are Now Operating in Days, Not Minutes

Task horizons have crossed from "draft a function" to "ship a small service," with the agent pausing only at strategic human checkpoints. The interesting consequence: code review is shifting from "did this PR work?" to "did this multi-day agent run obey the intent spec we gave it?"

That last phrase — intent spec — is the report's most actionable contribution. The reason engineers fully delegate so few tasks is that delegation requires persistent context: objective, outcomes, constraints, edge cases, verification. Without that, every agent run is improv. With it, the same task moves from the 0–20% delegated bucket to the routinely-shipped bucket.

Concrete Homework for PHP and Database Teams

Write an intent.md for every non-trivial PR. Five sections, max one page: Objective, Outcomes (the diff a reviewer would expect), Constraints (Laravel 12, MySQL 8, no new packages), Edge Cases, Verification (the exact php artisan test --filter=... line). Yes, this looks like 2010-era specs. The point is that a 2010-era spec was for humans who already had context; the 2026 spec is for an agent that will read it cold five times in one afternoon.

Stop one-shotting database migrations. Use a multi-agent pattern: a schema agent generates the migration, a query agent rewrites every affected Eloquent query, a security agent runs an authorization audit, a docs agent updates the API spec — orchestrated by a single Claude Code session reading a single intent.md.

Treat the 0–20% number as a roadmap, not a ceiling. Map your team's tasks on a "delegation maturity" axis. The tasks at 0% today (data migrations, auth changes, anything touching billing) are the ones whose intent specs you have not written yet — not the ones AI can't do.

My Take

Three years ago the question was "should we use AI to code?" Two years ago it was "which model is best?" The 2026 question is much less sexy: "are our specs good enough that a model we already pay for can finish the work?" Anthropic's report is, more than anything, a quiet reminder that the future winners of agentic coding aren't the teams with the fanciest models — they're the teams that finally learned to write down what they actually want.