Humans and AI agents, shipping together
We taught AI to code.
Now let’s teach it to collaborate.
Concurrent Delivery maps how humans and AI agents plan, coordinate and ship software together. They now change the same system at the same time, and making that work fit together is the open problem. Here are the challenges, the approaches people are trying, and what the evidence says so far.
The map
Where
the pressure builds up.
Agents working concurrently put pressure on every step from planning to remembering why. Open one for what practitioners say.
01 / PlanHow do you split work so parallel agents don’t keep colliding?4 quotes 02 / IsolateWhat does an agent need of its own to work without stepping on others?4 quotes 03 / CoordinateHow do agents avoid undoing each other’s work?7 quotes 04 / IntegrateHow do many parallel changes land without the queue becoming the bottleneck?4 quotes 05 / VerifyHow do you know the combined result of many agents’ changes is safe to ship?3 quotes 06 / RememberHow does the next session know what the last one decided?3 quotes
Latest from the research
What’s new.
What we’re learning.
The research moves as we find things. Here is what came in most recently.
Our notes
All our notesQuotes, links and videos
- Oct 14“Every agent would hit the same bug, fix that bug, and then overwrite each other's changes. Having 16 agents running didn't help because each…”Nicholas Carlini, Anthropic
- Oct 14“Worktrees reduce file collisions. They do not isolate services, ports, databases, secrets, or decisions.”Steve Kinney · 2 quotes
- Oct 26“In an early version of the “build a game” experiment in which agents built upon the same model all came online at the same time, 18 out of…”Anthropic research team · 2 quotes
Recently added tools
- Oct 24agentmashA pre-edit hook that tells a Claude Code or Codex agent when a teammate's agent has just changed the file it is about to edit.
- Oct 23ZitA git extension giving each agent a copy-on-write workspace instead of a worktree, with claims on files, symbols or Markdown sections and an accept step that refuses stale or interface-breaking changes before they land.
- Oct 22LayerFSOpen-source filesystem where each agent mounts a full writable workspace over one shared, checkpointed history.
Help build the map
Have something
worth sharing?
Know a tool, paper, post, talk, or tiny experiment we should be tracking? Send the link and a sentence about why it's interesting. Contributions are welcome—even if the idea is unfinished.
Contribute to the research ↗
One good link is enough to start.