The Reality of OpenAI's 'Ultra' Reasoning and Sub-Agent Bloat

The Misconception of 'Ultra' Reasoning

OpenAI's recently introduced 'Ultra' and 'Max' reasoning levels in tools like Codeex are frequently misunderstood. While marketed as advanced reasoning tiers, 'Ultra' is functionally a system prompt toggle that instructs the model to aggressively spawn sub-agents. Unlike a standard reasoning level, which adjusts the model's internal chain-of-thought, 'Ultra' forces the parent agent to delegate tasks to sub-agents, often recursively. This creates a massive, hidden cost: because the parent agent is set to 'Max' reasoning, every sub-agent it spawns inherits that same high-cost reasoning level, leading to exponential token consumption.

Sub-Agent Architectures: V1 vs. V2

Codeex currently utilizes two distinct sub-agent implementations. V1 acts as a simple dispatcher, hiring temporary helpers for specific, isolated tasks. V2, however, introduces a more complex 'org chart' structure featuring mailboxes, typed messaging between agents, and shared workspaces. While V2 allows for more sophisticated task decomposition, it is currently bloated and prone to context pollution. By default, V2 shares the entire conversation history with every sub-agent, which is computationally expensive and often unnecessary. The lack of a depth limit in V2, combined with these context-sharing defaults, is the primary driver behind users hitting their five-hour usage limits in minutes.

The Superiority of Programmatic Workflows

Theo argues that the current 'Ultra' implementation in Codeex is an inferior attempt to replicate the 'Workflows' found in Claude Code. Workflows allow for programmatic definition of agent behavior—using JavaScript to define phases, schemas, and specific model assignments for each step. By defining a clear pipeline (e.g., Review Phase -> Synthesis Phase), developers can control exactly which model is used at each stage, preventing the 'infinite recursion' of high-cost reasoning. This approach is more predictable, efficient, and provides better output quality than the current 'Ultra' toggle, which lacks granular control over sub-agent reasoning levels.

The Path Forward

For most users, the best advice is to avoid 'Ultra' until the implementation matures. OpenAI is actively iterating based on feedback, and the current 'Ultra' experience is largely a result of aggressive defaults rather than inherent model limitations. Developers should prioritize tools that allow for explicit configuration of sub-agent reasoning levels rather than relying on automated toggles that hide the underlying complexity and cost.