Codex smart agent

The smart agent is a costly, high-intelligence specialist for complex, high-judgment assignments. It uses GPT-5.6 Sol with high reasoning and should be dispatched sparingly, only when the quality benefit justifies the additional usage. Routine exploration, simple edits, and ordinary validation belong to the normal worker, explorer, or validator roles.

It is configured in .codex/agents/smart.toml and registered in .codex/config.toml under [agents.smart].

Capabilities

The role intentionally does not override sandbox_mode, approval_policy, mcp_servers, or tool settings. It therefore inherits the main/orchestrator agent's full shell, filesystem, MCP, approval, and network capabilities for the assigned task. Its only capability restriction is that it cannot spawn subagents.

No-subagent boundary

The role sets [agents] max_depth = 0, [features] multi_agent = false, and enable_fanout = false. These configuration guards prevent it from creating child-agent threads or fan-out work. Its developer instructions also explicitly forbid spawn_agent, spawn_agents_on_csv, delegation, and child agents.

The parent orchestrator can dispatch smart directly; the zero depth applies to further descendants created from the smart role, not to the parent's direct dispatch.

Configuration contract

Setting	Value
name	"smart"
model	"gpt-5.6-sol"
model_reasoning_effort	"high"
sandbox_mode	inherited from the orchestrator
approval_policy	inherited from the orchestrator
mcp_servers	inherited from the orchestrator
[agents].max_depth	0
[features].multi_agent	false
[features].enable_fanout	false
model_provider	(unset — inherit the orchestrator provider)

Use the role only for assignments with substantial ambiguity, architectural judgment, difficult debugging, or other work where the highest reasoning tier is warranted. The agent works directly and reports its result or blocker to the orchestrator without delegating.