One developer, two dozen agents, zero alignment
AI agents are built for one person working alone. Real work is done in teams, and that's the new bottleneck.
Apr 2026 · 3 min read
By Maggie Appleton on Blog
The short version
- Today's AI agents are “single-player”, but most real work is done by teams.
- When everyone's agents race ahead on their own plans, teams drift apart faster.
- Agreeing what to build is now the bottleneck, not building it.
- Plans, decisions and agent work should be shared with the whole team as they happen.
Who it's for: Teams where several people use AI on the same project, and the people who lead them.
Why it's genius
Maggie Appleton is a designer and researcher at GitHub Next, GitHub's research team. While everyone else celebrates one person running lots of agents, she points out the catch: software, like most work, is made by teams.
When each person's agents race ahead on their own private plan, a team can drift apart faster than ever. The usual checks, like tickets and code reviews, arrive too late, after the agents have already built the wrong thing at speed.
Her line sums it up: “Agreeing on what to build is the new bottleneck.” The talk also gives the first public look at Ace, a GitHub Next prototype where teammates and agents work in one shared space.
Try it
Before anyone sets an agent going, agree the plan together in one short shared note.
Share agent plans with the team, instead of keeping them in your own chat.
Invite designers and product people into the same conversations, so they see decisions as they happen.
Once a week, check what everyone's AI has built against the shared plan.
Before you build anything, write a one-page plan I can share with my team: the goal, what you'll build, what you won't, and the questions we need to agree on first.
Our take
Faster building makes bad plans more expensive
It sounds backwards, but AI makes alignment matter more, not less. When building took weeks, a vague plan had time to get fixed along the way. When agents build it in an afternoon, a vague plan becomes a finished wrong thing by teatime.
That's true well beyond software. A marketing team where everyone has their own AI churning out copy has the same problem: lots of output, pulling in different directions.
Spend the time you save on thinking
Her most generous idea is what to do with the time AI frees up. Instead of doing more things, do fewer things better: more research, more deliberate planning, more craft. That's the opposite of the “ship ten times as much” story, and we think she's right.
In practice that means planning together before the agents start, and keeping the plan going while they work, rather than treating it as a box ticked at the start.
The fix is mostly habits, for now
Ace is a research prototype, so you can't just install it. Until shared tools arrive, the fix is behaviour: write the plan down in one shared place, share agent plans instead of keeping them in your own chat, and bring designers and product people into the conversation early.
It also pairs neatly with Anthropic's research on meta-work. Anthropic shows managing AI eats into each person's time. Appleton shows it eats into the team's shared understanding too.
Where it shines
Any team where several people use AI on the same project. It names a problem most teams haven't spotted yet.
Where it breaks
The tooling she shows is a prototype. Today, most of this is down to habits, not software.
Where Ugenie helps
It's the team version of meta-work. Ugenie shows where everyone's AI work is up to, so drift gets spotted early.
Get Ugenie →Tools mentioned
GitHub Next Ace
A research prototype of a shared workspace where teammates and AI agents plan and build together.
- Cost
- Not publicly available; it's a research project.
- Our verdict
- Worth watching as a glimpse of where team tools are heading. Don't wait for it; start the shared-plan habit now.
Coding agents' plan modes
Most coding agents can write a plan before they act, which you approve first.
- Cost
- Included with the agent.
- Our verdict
- Use it, but share the plan with your team rather than approving it alone. That one change covers much of her advice.
Prices are rough and change often. Check before you buy.
Questions people ask
What does “zero alignment” mean?
It's when each person on a team uses AI agents to race ahead on their own plan, so the team stops agreeing on what it's building. Maggie Appleton argues this is now the main bottleneck.
Is this only a problem for software teams?
No. Any team where several people use AI to produce work, such as marketing, sales or operations, can end up with lots of output pulling in different directions.
What is Ace?
A GitHub Next research prototype that puts teammates and AI agents in one shared workspace. It isn't publicly available yet.
All credit to Maggie Appleton. We share things we think are genius and add our own take. Watch or read the original for the full thing.