Optimize for generation.
Agents finish bounded tasks quickly. The team gets more work than it can review and own.
Attention Deficit · episode 003 · visual 03HumanLayer · builder retrospective
Why software factories fail
The factory got faster.The queue did not.
Agents make implementation cheap. The team still has to review, understand, and own the work.
The opening chart starts with generated work outrunning review. The main chart lets you move decisions earlier.
The review-capacity model
Decide earlier. Keep the backlog small.
The chart opens with fast generation and slow review. Advance the scrubber or adjust both rates.
Reduce the backlog.
Small vertical slices keep the backlog small and keep review close to generation.
Decide before the diff.
Product intent, architecture, and program design constrain the agent before it adds to the backlog.
Match generation to review.
The team still uses agents. It chooses a rate that people can review and redirects work while changes remain cheap.
The backlog grows faster than the team can review it.
generatedreviewedbacklog
after generationduring designbefore generation
backlog at horizonlarge
late reworklikely
team understandingunder pressure
01 · THE SHIFT
Generation got faster. Review did not.
An agent builds the ticket. The team still has to review, test, understand, and own the generated work.
stage
human-built
agentic
build a change
hours or days
minutes or hours
review a change
hours or days
hours or days
architecture
built during work
rebuilt during review
02 · TIME SCALE
Tests grade the change now. Architecture sends the bill later.
Automated checks can catch a local defect in seconds. Maintainability appears when someone changes the system again, often weeks, months, or years later.
SECONDSTests catch local defects.
HOURSReview checks the change.
DAYSIntegration exposes system fit.
WEEKSArchitectural erosion appears.
MONTHS+Ownership cost becomes visible.
03 · RESPONSE
Make the big decisions while change is cheap.
HumanLayer kept the agents. The team moved its decisions back into product intent, architecture, program design, and small slices.
Product intent
Name the user, outcome, and reason.
Stops the agent from solving the wrong problem quickly.
System architecture
Set boundaries and ownership before implementation.
Stops local correctness from eroding the whole system.
Program design
Choose interfaces, data flow, and failure behavior.
Makes the diff easier to review because the team already understands the structure.
Small vertical slices
Generate less work before the next decision.
Keeps the backlog small and redirection cheap.
Review and redirect
Review while the team still shares the context.
Turns review into steering, not archaeology.
04 · OWNERSHIP
Delegate the work. Keep the decisions that shape the future.
Select a decision below. The useful split is not human versus agent. It is durable judgment versus bounded execution.
Keep human judgment here
What should exist?
What is the root cause?
Who owns the result?
How do we know it will last?
“The lights-off factory did not remove humans. It moved them to review and flooded the backlog.”
The useful gain is faster implementation inside a system the team still understands.