17 lines
1.1 KiB
Markdown
17 lines
1.1 KiB
Markdown
---
|
|
id: 23bb2301-21ee-4a02-8e4b-64e4eca3973b
|
|
type: decision
|
|
title: "Wave-based agent delegation avoids file contention"
|
|
tags: [cognitive-memory, agent-delegation, pattern, architecture]
|
|
importance: 0.8
|
|
confidence: 0.8
|
|
created: "2026-02-13T18:03:08.417460+00:00"
|
|
updated: "2026-02-13T18:03:08.417460+00:00"
|
|
---
|
|
|
|
When delegating multiple tasks to agents that all modify the same file, run them in waves instead of all at once. Wave 1 (3 agents): additive changes that touch different sections - worked cleanly. Wave 2 (2 agents): one finished first and the second hit file contention errors until the first was done. Solution: tell the blocked agent to wait and retry after the other finishes.
|
|
|
|
Key pattern: group agents by whether their changes overlap. Additive-only changes (new methods, new constants, new CLI subcommands) can run in parallel. Changes that modify existing methods should be sequential.
|
|
|
|
Also confirmed from Phase 1: always check that get() returns new fields added by agents. They consistently add fields to store/write but forget to update get() return dict.
|