Codex’s multi-agent design still forks full context for every subagent
bdsqlsz · x · 2026-07-23
A user says Codex still has a serious multi-agent design issue: the main thread cannot access spawned agent metadata, so subagents effectively operate at the same level as the parent. The screenshot points to forkturns defaulting to all, which means every subagent gets a full context fork and then context must be compacted again, suggesting the setting should be either disabled or reduced to only a couple of turns.
Related event: Codex Multi-Agent Architecture Criticized for Context and Metadata Flaws(2 posts)→
More from coding & agent
- Agent orchestrators are shifting toward dynamic workflows and meta-harnesses — omarsar0 · 2026-07-24
- Dynamic agent workflows unify routing, loops, graphs, and multi-backend orchestration — omarsar0 · 2026-07-24
- LangChain points to a fireside chat on common Deep Agents applications — LangChain · 2026-07-24
- Uncle Bob says coding agents are only trustworthy when tests and QA box them in — jdjohnson · 2026-07-24
- Grok 4.5 in Cursor feels close to planner-and-executor territory — brandon_galang · 2026-07-24
- SubText is an open-source file format that keeps sentence-level notes with prose — HamelHusain · 2026-07-24