Your Team is Doing AI Wrong
I bet your team is going about AI agents all wrong. And it’s not because of which tools or models used.
Quick → How many agent sessions do you have open right now? Three? Five? More?
Cool, so → How many of those tasks will you actually finish and ship today?
Did you ship tho?
The big promise of AI agents in software development is increased velocity and enablement. PMs are prototyping, designers are pushing PRs, and engineers are shipping features in hours instead of weeks.
But there’s a version of this playing out in most orgs where tool adoption and token spend is up, but if you look closer you’ll see that velocity and cycle time haven’t actually budged.
So, what gives?
Agents ≠ higher velocity automatically
Agents, when wielded without discipline, exploit our temptation to multitask. With some tools it’s as simple as a single click to assign an agent to a task, and before you know it you’ve kicked off six parallel workstreams before lunch. But to increase velocity that work needs to get “over the line” (actually shipped!).
These separate sessions all need your human input and judgment (reviewing what the agent built, deciding if it’s right, editing it, shipping it), and with six sessions running in parallel, you’ve created six queues, each waiting for your judgment, each demanding context-switching overhead when you return to them.
WIP is WIP
In our haste to move faster, it looks like we’re forgetting some of our SDLC foundations – like the benefits of limiting work-in-progress. Remember Kanban? Remember “stop starting, start finishing”?
The insight behind WIP limits isn’t complicated: the more things you’re working on at once, the longer everything takes. Every time you put down some partially completed work you’re expanding the team’s overhead (context-switching, blocking dependencies, the cognitive load of keeping a dozen things in a half-baked state).
So, if you’re often ending the day with more WIP than you started with, are you and your agents actually productive? Maybe, instead, you’ve just created a bunch of in-flight work.
So, multiple agents running is bad?
Multiple agents running at once isn’t inherently a problem (I have 3 sessions going right now).
But finishing and shipping even just one thing is more valuable than any amount of open agent sessions on your local machine.
I said what I said. 🤷♀️
Remember that shipping is the point
Most organizations pushing on “use AI to increase your output” are still optimizing for the wrong thing. Token spend and active seat count are signals of adoption (ie: who is using it), but they are not signals of how efficacious using these tools are on your workflows. Is your team actually shipping faster than before?
Mandates to “use AI to go faster” without offering up some best practices for an effective agent-native workflow is how you end up with a team that feels busy, looks productive on the surface, and wonders why work keeps carrying over to the next sprint.
These best practices aren’t new: limit what’s in flight, finish what you start (or close it), and remember what your definition-of-done is (it’s probably something like “shipped to users”).
The teams that are genuinely moving faster with AI are the ones who’ve figured out how to use agents to complete, review, and ship a unit of work faster, not just to accelerate starting tasks. They are using agents as an accelerant for an already strong, disciplined development workflow and not as a replacement for discipline.
This is your reminder to keep our software engineering fundamentals top of mind in this new era of building.
Korey is a workspace for agents and people to build software in a words-first (vs code-first) world. Your team shapes the spec and your favorite coding agents do the work underneath. Tasks and code stay in sync as the spec changes.
Built by Shortcut, the project management tool where humans and agents collaborate to build great software, used by tens of thousands of teams.