Concepts/Definition

Application context: two agents can overwrite each other without touching one file

Two coding agents work in separate worktrees on different services. Both allocate REQ-417 from the same identifier table, or one changes a client while the other changes a server under incompatible assumptions about the public API. Git reports no file conflict; the business writes still conflict.

Definition

A Write-Conflict Domain is a set of objects that cannot be mutated concurrently without coordination. It may contain files, identifier spaces, shared configuration, database rows, test environments, API contracts, deployment slots, cloud quotas, or external business resources. The domain is defined by what the tasks jointly change, not only by repository paths.

Engineering mechanism

Before scheduling, each task declares an expected read/write set and resource keys such as api:customer-v2, db:tenant-17/payroll, or env:uat-3. The scheduler chooses serialization, sharding, leases, or merge rules. Discovery of a new shared object expands the domain and triggers rescheduling instead of allowing the agent to continue under a false isolation assumption.

Boundary

Perfect static inference of every write set is rarely practical. Unknown high-risk objects should begin serially and move to parallel execution only after a partition rule is known. Worktrees remain useful for source files but cannot replace conflict checks over domain objects, external resources, and versioned contracts.