Flow is broken

Flow is broken. Csikszentmihalyi’s flow depends on uninterrupted immersion and is sustained by a sense of authorship.

Mihaly Csikszentmihalyi described flow as the state where attention and action merge. Time seems to disappear. The research after him names three conditions: clear goals, immediate feedback and a balance between challenge and skill. Among its features are the merging of action and awareness and a sense of control. “Authorship” is not a word from his list. The closest match is that sense of control, which psychologists study in the lab as the sense of agency.

§The waiting

1. Uninterrupted immersion is completely broken when waiting for LLMs/Agents to do their job. People look for hacks like keeping several parallel sessions which brings only a small sense of satisfaction.

Immediate feedback goes first. Write a line of code and it compiles or it doesn’t. The loop is tight, so you never leave the task. With an agent, you write the instruction and wait. Thirty seconds, a few minutes, sometimes twenty for a larger job. Each wait is an exit to email or another tab. Coming back, you reload the problem before you can judge the output. Parallel sessions fill the gaps. While one agent works, you steer another. One deep task becomes a rotation of short check-ins.

§Whose thought is it?

2. While for authorship, things are more subtle. Prompting introduces a persistent uncertainty about whose thought it really is. When you instruct, you become a cause of a cause. Our sense of competence and self-efficacy is tied to direct causal contact with outcomes.

When you instruct, you become a cause of a cause.

The “cause of a cause” can be measured. When people act voluntarily, the gap between action and outcome feels shorter than it is. Patrick Haggard and colleagues first reported this in 2002 as intentional binding. In 2016, Emilie Caspar and colleagues had people press a key on an experimenter’s order. They showed a reduced binding effect. Their brain responses looked closer to passive movement than to intentional action. In 2026, Edwards and Obhi tested instructions from an on-screen chatbot. Binding was strongest under free choice and significantly reduced when the chatbot directed the action. These are simple lab tasks, not agentic work. They still point the same way.

Self-efficacy is Albert Bandura’s term. He argued that mastery experiences are its most influential source, because they are the most direct evidence. A pull request an agent wrote and you approved is evidence too. It is one step further removed.

I’ve been turning this over for weeks and still no answer. What’s your take?