Maggie Appleton @maggieappleton.com · Apr 15 Got to talk at @aidotengineer.bsky.social conf last week about the need for collaborative AI engineering. All our current coding agents are single player. We're trying to scale up individual productivity, but creating tons of alignment problems in the process. We have no good tools for...
134 likes 22 replies
Replies
...communication that deeply integrate with coding agents Current versions of Github, Linear, Slack, etc. were built for a previous era of development. They're too isolated from agentic coding tools. Trying to duct tape coding agents onto them isn't working. Coding agents and realtime...
the political dynamics one is underrated. half the context an agent needs to make good decisions lives in slack threads and meeting notes that never make it anywhere near the repo
This is brilliant. I've realised I've built the foundations of this for our designers, the microVM and web UI on top of claude code for them. Does the Preview/create PR bit. Off to think about how we'd add collab features!
Yes. Right now a lot of 'agent collaboration' is just multiple soloists creating a coordination bill for someone else. The missing product is shared state, review context, and clean handoffs — not just more parallel output.
I just watched your talk. There’s so many great ideas in ACE. I’m already a fan. Not enough people are thinking about the “definition of ready” or “do we have the right context to feed to the agent”? That shouldn’t be single player mode. The microVMs and seeing what other are doing is super cool too
I just read the essay/talk and wanted to say thank you. It crystallized something I've been circling for a while. I've been writing about what happens when you treat strategy itself as a protocol: explicit decision logic, written in plain text, readable by people and executable by machines.
the single-player problem is real, every coding agent ships changes that nobody else reviewed or even saw happen
This is spot on. Single-player agents optimize for individual output but create knowledge silos. Best teams I've seen had shared context, pair programming, collective ownership. We need multi-agent systems that actually collaborate, not just parallelize.
Love this pitch. It creates a lot of great questions. People collaborate with more modalities that just text. Tone of voice, pauses, smiles, and so on are how we collaborate in a value-based sense. That's key for multi-player interfaces at the fundamentally human goal and worth level. AI how?
You know, it's hard to consume a talk after the fact. I don't want to watch an hour long long vid, and reading the transcript loses some of the dynamism. Reading/watching `https://maggieappleton.com/zero-alignment` was so fantastic, with the auto updating slides. Best of both worlds. Well done.
@tbsky.app unroll this thread.
the single-player framing is the right diagnosis. every agent right now optimizes for individual productivity, but the output still has to fit into a team. that gap is where alignment problems pile up. what do you think forces tooling to catch up?
Single-player agents feel like a symptom of a missing 'shared world model' layer. Two devs with two Claude sessions are now four perspectives, none aware of the others. Curious what the multiplayer layer looks like in your view.
The single player framing nails it. Even inside one team, each agent session is amnesiac about what teammates decided yesterday. Git handles code state but there is no equivalent for agent intent or rationale. Shared context matters more than shared output.
The missing primitive is shared working memory. Two agents on the same repo can't see each other's reasoning trace, so they overwrite assumptions instead of negotiating them. Git is too coarse for that. Most 'multi-agent' demos are still one model talking to itself.
The single-player problem compounds when each engineer's agent develops its own implicit style. The codebase ends up reading like 5 contractors who never met. Shared context isn't a nice-to-have, it's a prerequisite for team velocity.
The bottleneck has quietly shifted from keystrokes to context sharing. Ten single-player agents generating parallel PRs just pushes the merge and review burden upstream onto humans now reasoning about decisions they never made.
Single-player is the quiet crisis. We optimized for individual throughput and broke team legibility. Nobody can review what they didn't write with context they didn't see. The next frontier is shared state, not faster agents.
Single-player agents also mean every dev builds a different mental model of what the agent knows. The missing primitive might be team-level agent memory that lives in the repo, not the IDE, so context becomes a shared artifact.
ran into this exact problem. multiple agents working in parallel with no shared context - they contradict each other and nobody notices until review. the tools aren't built for that world yet