Agent Coordination for Parallel AI Coding Agents
How Araptus runs many AI coding agents on one codebase
Parallel AI coding agents collide even in separate worktrees. How Araptus coordinates a fleet of agents on one codebase, and what the measurements show.
Running parallel AI coding agents is easy to start and hard to keep safe. Give each agent its own git worktree and they stop overwriting each other's files. They don't stop building the same feature twice, trusting a stale status, or colliding on the one machine they all share.
Araptus builds on a coordination layer made for exactly that problem. It is internal, and it isn't for sale. This page is what running a fleet of agents under it has taught us, with the numbers and the dates they were measured.
Why Parallel AI Coding Agents Collide
Worktrees isolate files. They do not isolate decisions.
Everything a team of agents shares is still shared after you split their files apart: the history, the database, the migration sequence, the build machine, and the main branch every piece of work has to land on. Isolation doesn't remove those collisions. It makes them quieter. Two agents in separate workspaces can each write a clean, self-consistent version of the same feature, and nobody notices until one copy is thrown away at merge.
We wrote up the failures in detail in why parallel AI coding agents collide. The short version: in none of them did an agent misbehave. Each one acted correctly on information that was wrong and couldn't say so.
What a Coordination Layer Does
A coordination layer gives every agent on a codebase one shared record of the work, so each can see what the others are doing before it starts, not after it merges.
Who is doing what
Before an agent starts, it can see whether someone else is already doing that work.
What has been learned
A finding one agent pays for is available to the next, instead of being rediscovered in every session.
What actually shipped
"Finished" and "merged" are recorded as different things, by different roles, because they answer different questions.
Your agents will coordinate whether you build them a way to or not. The only real choice is whether you can see it.
Four Rules From Running an Agent Fleet
- An instrument must be able to say "I don't know". The worst reading is a confident one that is wrong. A status that can't express its own uncertainty will be read as permission.
- Structure beats instruction. Telling agents about a failure mode doesn't stop it. We measured agents repeating a failure they had personally written warnings about. What works is an arrangement where the wrong move can't be made.
- Never ship a field only a well-behaved agent keeps true. If staying accurate depends on an agent remembering to update it, it will go stale. Make accuracy a byproduct of doing the work.
- One role owns merging. Saying work is done and saying it is merged are separate statements, made by separate roles. Mixing them up is how a finished task gets read as a shipped one.
What Eleven Days on One Codebase Looked Like
Five agents, one codebase, eleven days, as measured on 8 September 2026. Each figure is a count of distinct records, not of logged events.
324
tasks finished
261
landed on the shared branch
904
written findings
1,674
messages between agents
After the merge role was written down, merges by anyone other than the role holder fell from about a quarter to two per cent. We report that as a correlation, not a result: the same period changed many things at once.
Independent Research Reached the Same Conclusion
Two research groups published on coordinating coding agents in 2026, and both found that coordination, not parallelism, is what matters.
- Before the Pull Request measured agents sharing work with a shared coordination record: the share of work that merely redid a teammate's task fell from 78% to zero, and useful throughput more than tripled.
- AgentRoom tested concurrent coding agents in a shared workspace and concluded: “Coordination, not parallelism or CRDT-merge, bears the load.”
What This Means If You Build for Clients
Every platform Araptus ships is built this way, starting with Highlander. One engineer runs a fleet of agents, and the coordination layer is what keeps that fleet from quietly undoing its own work.
If you're a developer, an IT or security firm, or an agency, you can build on those platforms for your own clients, partner with us on a project, or resell them under your brand. The same thinking runs through the rest of our writing: why the most powerful AI tool is a folder and how I build, terminal first.
The bottom line
Isolation gives each agent a safe place to stand. Coordination tells it where everyone else is standing.
Build with us on platforms made this way.
Sources
- Sarkar — Before the Pull Request: Mining Multi-Agent Coordination (arXiv, June 2026)
- Cho & Lee — AgentRoom: Concurrent Multi-Agent Coding in a CRDT-Backed Shared Workspace (arXiv, August 2026)
- Fleet figures are first-party, read from our own coordination records as of 8 September 2026. The method and caveats are in the full write-up.
