AI made starting work almost free. It did not make finishing any cheaper. The only thing that counts now is what actually gets finished and shipped.
I learned this running coding agents on my own work. Three issues at a time felt calm. I could read every plan, steer every session and review every result while the context was still in my head. Then I tried eight. Eight got out of hand fast. Sessions sat waiting on me, I lost track of which branch was blocked on what, and the review queue grew faster than I could clear it. The agents were not the problem. I was the finishing step, and I had opened more than I could finish.
The same thing is happening across knowledge work. A first draft of a memo, a market analysis, a set of campaign variants or a contract redline now takes minutes to start. Every one of those still needs a person to read it, decide on it, fix it and send it. Count honestly and most teams carry far more open work than anyone would pick on purpose. Nobody chose that number. It is what happens when starting costs nothing. Little's law says the time an item spends in the system equals the work in progress divided by throughput, so if finishing capacity stays flat while open work climbs, everything slows down, including the work that matters most.[1] It is the same wall I described in why merging, not generation, is the constraint, seen from the planning side.
So stop measuring activity. Drafts generated, agents running and tasks opened all look like progress, and none of it is until it ships. The count that matters is finished work that reached a customer, a decision or production. That is also why counting your agent mechanics beats counting your agents.
Then set the limit on purpose. Pick an explicit concurrency number for each team, the way you would size a worker pool, and let new work wait until a slot opens. When finished work clears faster than new work arrives, raise the number. When items start aging in review, lower it. This week, count what your team actually has open, pick a number below it, and see what ships a month from now.[2]
Key takeaways
- AI made starting work nearly free, so most teams now carry far more in flight than anyone chose.
- Finishing still costs the same, because every open item needs a person to review it, decide on it and hold its context.
- Count only what ships, and set an explicit limit on how much work can be open at once.
FAQ
Does this apply outside software teams?
Yes. Drafts, analyses, campaign variants and contract redlines all got cheap to start, and every one of them still needs someone to finish and approve it.
How do you pick the right amount of work in flight?
Start low enough that every open item gets real attention, then raise the limit only when finished work clears faster than new work arrives.
Related Essays
Generation Is Solved. Merging Is Not
Coding agents made producing pull requests nearly free. Now the backlog lives in review, and the teams that win will engineer the review layer, not just generation.
Review Needs a Computer, Not a Diff
Agents have made code generation cheap and review the bottleneck. The fix is giving every reviewer a live machine with the change already running, not a diff in a browser tab.
Agents Start Broken and That Is the Point
New agents should deploy in supervised mode by default because no agent is production-ready on day zero.