The environment changes what people delegate.
What Anthropic’s Cadences report suggests about tools, delegation and learning—and what it cannot prove.
An assistant in a conversation and an agent working inside a task environment can be powered by the same model while supporting very different ways of working. Anthropic’s June 2026 Economic Index examines that distinction. For an operator, it is a reason to look at the whole workflow rather than only the model.
The delegation gap survives a same-model comparison
Claude Code minus chat / Cowork · points on a 1–5 autonomy scale
Reported benefits extend beyond speed
Share of linked survey respondents reporting gains · percent
What was observed
Claude Code sessions averaged 0.37 points more autonomy than chat and Cowork on a 1–5 scale. Among Sonnet sessions, the difference remained 0.26 points. The product environment therefore appears relevant alongside model choice, although the comparison is observational.
In the linked survey, 86% reported gains in speed, 82% in scope, and 69% in quality; 68% reported learning more. The sample included approximately 9,700 respondents and was not representative of the general population.
Our interpretation: teach delegation, not just prompting
Prompt-writing is one skill. Delegating a piece of work involves several more: defining completion, choosing the source of truth, deciding which actions require approval, and identifying when to stop and ask a person.
A useful training session begins with a task from the team’s own queue. One person explains the business judgment. Another checks whether the draft respects the current sources and rules. Together they refine the brief and preserve the parts that will matter next time.
The goal is not to make every employee a software developer. It is to help them recognize what can be delegated and what they must still inspect.
A practical test for your company
Choose a task where the current process involves collecting context from several places. Prepare the same request in two settings: an isolated conversation, and a permissioned workflow with the relevant documents and tools already available.
Compare the resulting work, the number of manual handovers, and the review effort. Keep the person and the business question as consistent as practical. Treat this as a local pilot, not as a scientific replication of the report.
If the connected workflow saves effort but produces a less trustworthy result, it is not ready. Improve the sources or narrow the authority before expanding it.
Learning is a responsibility, not an automatic outcome
Surveyed learning and productivity are self-assessments. People who are enthusiastic about AI may both delegate more and report better outcomes. The findings do not prove that automation causes learning, rule out skill erosion, or establish a universal productivity gain.
Make review part of the working habit. Ask the team to explain why an output is sound, where its sources came from, and which exception still needs judgment. That is a stronger signal of capability than accepting the answer quickly.
Take it into the next conversation.
When assessing an AI rollout, ask what the employee learned to delegate, what context the system can access, and how the employee checks the finished work.
Sources & further reading
Anthropic: Anthropic Economic Index report: Cadences June 2026 report. Observational usage comparisons and a self-selected linked survey sample; not a causal experiment.
Source pages checked 19 October 2026. Our examples and recommendations are editorial interpretation, not findings from those publications.Have a workflow in mind?
Let’s talk about the work