Detailed Analysis
A Reddit user reports what appears to be a significant discrepancy between instructed behavior and actual execution in Claude Code, Anthropic's agentic coding tool. The user configured the primary model as "Fable" and explicitly instructed it to reserve that model for high-level strategy and judgment calls, while delegating routine tasks like data fetching to Sonnet or Opus worker agents. According to the post, Claude Code initially acknowledged this instruction, confirming it would spawn Sonnet/Opus subagents for the delegated work. However, when the dynamic workflow actually executed, it reportedly spawned a swarm of Fable-only agents instead, ignoring the explicit model-routing instructions. The consequence was severe: usage quota went from 0% to 100% in roughly fourteen minutes on a Max Plan subscription, effectively burning through a premium allotment due to what the user believes is a bug in how the orchestration layer selects models for subagents.
This kind of report matters because it touches on a core reliability concern in agentic AI systems: the gap between an LLM's stated intentions and its actual tool-use or delegation behavior. Claude Code's subagent architecture is designed to let a "manager" model coordinate specialized workers, ideally allocating expensive, high-capability models only where genuinely necessary and cheaper or faster models for rote tasks. When that routing logic fails silently, users lose the cost and quota efficiency that the multi-agent design is supposed to provide. For subscribers on metered or capped plans, this isn't a minor inconvenience but a direct financial and productivity issue, especially if the discrepancy between acknowledged plan and executed plan is systematic rather than a one-off fluke. The user's request for confirmation from "Anthropic employees" underscores a broader frustration common among power users of agentic coding tools: the primary channel for surfacing bugs is often community forums rather than a transparent, responsive support pipeline from the vendor.
More broadly, this incident is emblematic of growing pains in the industry-wide shift toward multi-agent, multi-model orchestration systems. As companies like Anthropic, OpenAI, and Google push agentic frameworks that dynamically spawn subagents, route tasks across models of varying cost and capability, and manage long-running autonomous workflows, the reliability of the orchestration layer becomes as important as the underlying model's raw intelligence. Bugs in this layer are particularly insidious because they can be invisible until a user notices unusual costs or context switches, as happened here. This also raises questions about observability and auditability in agentic systems: users need reliable ways to verify which models were actually invoked, not just what the system claims it will do, especially as these tools take on more autonomy in software engineering pipelines.
Finally, the episode reflects a recurring tension in fast-moving AI product development between rapid feature rollout and mature quality assurance. New model routing features and dynamic workflows are being shipped quickly to compete in the agentic coding space, but as this report suggests, edge cases in subagent model selection may not be fully hardened before reaching production. For enterprise and power users relying on predictable cost structures, resolving this class of bug is important not just for user trust but for the broader legitimacy of subscription-based agentic AI tools, where unpredictable resource consumption undermines the value proposition of tiered plans like Max.
Read original article →