Safe and durable collaboration · chapter 13 of 17
One task, one owner: agents and subagents
Multiple agents can work in parallel when roles are bounded. The main agent retains user intent, production decisions and final verification.
13.1 · plain explanation
How do we divide work without collisions, duplicate ownership or unresolved production changes?
Multiple agents can work in parallel when roles are bounded. The main agent retains user intent, production decisions and final verification.
Important: Parallelism helps when work can be isolated. If two agents must change the same file or production resource, serialize the work.
13.2 · building blocks
Three concepts to keep separate
Main agent
Owns scope, shared state, production, synthesis and the response to the user.
Worker agent
Receives one concrete objective, exact files, permitted actions and a stop condition.
Verifier
Reviews the diff and evidence independently but does not inherit decision authority.
13.3 · method
A practical workflow
- 1Split work by file or system ownership.
- 2Give every agent a written assignment packet.
- 3Allow only one writer per file or shared state.
- 4Have the main agent verify reports before merging or production.
13.4 · concrete example
Three subagents
Three subagents
How do we divide work without collisions, duplicate ownership or unresolved production changes?
One writes web content, one maps firmware read-only and one checks server status read-only.
The root AGENTS.md provides shared rules, but the coordinator, not the folder, assigns roles, authority and dependencies.
13.5 · try it yourself
Exercise
Divide an imaginary job into a main agent, builder and verifier. Write the objective, file boundary, forbidden actions and report format for each.
13.6 · checklist
Before you continue
- Does each subtask have one owner?
- Are write surfaces collision-free?
- Does production remain under main-agent control?
- Are agent reports treated as evidence rather than authority?
13.7 · sources
Sources
Reviewed 23 September 2026. Official sources support product-specific claims; the practical contract, memory and verification patterns are Finnandre project methods.