The Delegation Gap: Why 60% AI-Assisted Doesn't Mean 60% Delegated

Anthropic's 2026 Agentic Coding Trends Report puts a number on something most engineering leads have already felt. Developers now use AI in roughly 60% of their work, but report being able to fully delegate only up to 20% of tasks. That gap isn't a maturity curve waiting to close on its own. It's a structural line between "AI touched this" and "AI can be trusted alone with this." The report calls it the collaboration paradox. I'd call it the delegation gap, because the interesting question isn't that the gap exists. It's what decides which side of it a task lands on.

Source: Anthropic, 2026 Agentic Coding Trends Report (PDF).

Why this matters now

Every roadmap conversation in 2026 has some version of "just let the agents handle more of it." That instruction means nothing until you know what actually moves a task from assisted to delegated. Get it wrong and you either bottleneck your team with unnecessary review, or you ship something nobody actually verified.

The gap, in one picture

AI-assisted work ~60% Fully delegated tasks 0-20% Same engineers, same tools. Most work is touched by AI, very little is handed off entirely.

What actually decides delegation

The report's own explanation is the useful part. Engineers delegate tasks that are easily verifiable, where they can "relatively easily sniff-check on correctness," or that are simply low-stakes. Design judgment, organizational context, and taste are the opposite of checkable. There's no sniff test for "is this the right thing to build," so those tasks stay on a human's plate, or get worked through side by side with AI rather than handed off. That's a verifiability and stakes axis, not a task-complexity axis. A genuinely hard but checkable task, like whether an implementation matches a reference method, can fully delegate. A simple but judgment-heavy task, like whether you should build the thing at all, never does.

What this looks like in practice

These are Anthropic's own customer examples from the report, so treat them as self-reported wins rather than independent verification. Still useful for spotting the pattern.

Company What happened Which side of the gap
Rakuten Claude Code implemented an activation vector extraction method inside vLLM (12.5M lines, multiple languages) in 7 hours of autonomous work, matching the reference method to 99.9% numerical accuracy. Fully delegated. The task had a precise, checkable success criterion set in advance.
CRED Doubled execution speed across their fintech dev lifecycle. Engineers didn't disappear from the process, they moved into higher-value review and direction work. AI-assisted, not delegated. The speed came from collaboration, not hand-off.
TELUS Shipped engineering code 30% faster and built 13,000+ custom AI solutions internally. Anthropic's broader research puts a number on why: roughly 27% of what AI touches is net-new work, small quality-of-life fixes that used to lose the prioritization fight because they weren't worth a sprint. Mostly new, low-stakes work. The kind that was always delegatable, just never prioritized.
Fountain Fountain's hierarchical multi-agent setup, an orchestrator coordinating sub-agents for screening, document generation, and sentiment analysis, cut one logistics customer's time to fully staff a fulfillment center from over a week to under 72 hours. Delegated, but only after the task was broken down into narrow, independently verifiable sub-tasks.

The lead's job in this gap

None of these examples got easier because someone decided to "trust the agent more." They got easier because someone did the unglamorous work of narrowing a task until it had a checkable success criterion, or explicitly decided it was low-stakes enough to skip a human check. That's spec work, not delegation work. It's the same discipline as writing the spec before the code, just applied to deciding what an agent is allowed to own outright.

Before assigning a task to an agent unsupervised, answer one question. How will I know if it's wrong, and how much does it cost me if I don't catch it? If you can't answer both quickly, the task stays on the 60% side of the gap, assisted rather than delegated, until you can.