Restricting Multi-Agent V2 to a maximum of 4 concurrent threads, setting Terra with high reasoning effort as the sub-agent default, and delegating the "when to delegate" logic to AGENTS.md or the root prompt serves as a reusable starting point for quota and capacity control.
Suitable tasks: Codex CLI Multi-Agent V2, parallelizable and bounded engineering subtasks.
Unsuitable tasks: Serial modifications heavily dependent on shared state, production tasks without concurrency or write isolation configured.
Applicable model versions: GPT-5.6 Terra; the original post also notes that model and reasoning levels can be overridden on a per-bounded-task basis.
Applicable clients, agents, or APIs: Codex CLI 0.145.0 and versions supporting Multi-Agent V2 (the original post cites Codex Releases).
Recommended reasoning tiers and parameters: Sub-agent default high; up to 4 concurrent threads; max_depth = 1 retained solely for V1 compatibility.
The original post disclosed configuration intent and fields rather than a complete config.toml. It can be mapped into the following checklist based on the current Codex version:
Multi-Agent V2: Enabled
Expose per-generation model override and reasoning override
Max concurrent threads: 4
Sub-agent default model: GPT-5.6 Terra
Sub-agent default reasoning: high
Preserve interrupt context
max_depth = 1 (V1 compatibility only, not used as V2 delegation depth control)
Root Agent / AGENTS.md determines model and reasoning overrides during delegationEnable V2 in a Codex CLI version that supports Multi-Agent V2.
Set the default sub-agent model to Terra, default effort to high, and set the concurrency limit to 4.
Do not rely on max_depth to control V2 delegation depth; clearly define delegation conditions in AGENTS.md or the root prompt.
Override the model/reasoning level as needed for bounded tasks, and preserve interrupt context.
Monitor quotas, failure retries, concurrency conflicts, and subtask artifacts before adjusting the limits.
The original post did not disclose the full configuration file, hardware, task suite, run counts, or quota consumption, so these settings cannot be claimed as an optimal solution.
"Controlling capacity via configuration and delegating timing via instructions" reflects the author's personal experience rather than an official Codex guarantee.
Specific field names and available models should be based on the current CLI version; avoid blindly copying non-existent configuration keys.
GPT-5.6 Terra