Cole Medin reflects on the attention cost of parallel coding agents in this 16-minute video, then describes the workflow he is building to reduce supervision.

Faster output, more fragmented attention

  • Multiple agents can work simultaneously while their human supervisor keeps switching problems.
  • Medin misses four-hour coding sessions with music; his current day alternates among agents, content and course planning.
  • Waiting for one agent creates pressure to open another, turning productivity gains into a higher expected workload.
  • These are personal observations and anecdotes, with no measured productivity or attention study.

Spend the next improvement on autonomy

  • A harness is the software around an agent that runs its tools, instructions and checks.
  • Make that workflow handle longer tasks before adding more simultaneous conversations.
  • Specify intent up front, then let the agent plan, implement and validate against project-specific requirements.
  • Batch decisions and review completed work in staging, a test version of the application.
  • A generic /goal command still needs the working conventions and validation criteria that make delegation trustworthy.

The software-factory approach

  • Medin’s open-source factory turns release planning into linked issues and supervised automated delivery.
  • Keep human judgment concentrated at planning and acceptance, with more of the middle running in the background.
  • Building the system costs output today; Medin expects that investment to reduce repeated interruptions later.
  • He describes future day-long autonomy and one main planning conversation as a direction, not a reliability result established here.

Faster replies could help—but expectations still matter

  • Fast responses in the main conversation could reduce the temptation to switch while waiting.
  • Token-generation speed alone does not establish fast or correct end-to-end work.
  • Medin acknowledges the counterargument: people may simply juggle several factories and recreate the same overload.
  • His account leaves that behavioral problem open; greater automation does not guarantee less work.

“Our cognitive bandwidth doesn’t parallelize.”

Disclosure: the video includes a paid Revolut Business segment about agent spending controls. Its product assurances are the sponsor demonstration’s claims, not an independent security evaluation.