I gave a research task within Claude code app, it used over 100 agents. Was using max effort, and not ultra code, and used over 2M tokens in 19Mins. Is this Ultra code? Can max change to ultra if there is a need for it?

Detailed Analysis

A user of Anthropic's Claude Code platform reported a striking episode of autonomous multi-agent behavior during a research task, in which the system spawned over 100 concurrent agents and consumed more than 2 million tokens within a 19-minute window — all while operating under the "Max" effort setting rather than the higher "Ultra" tier. The observation raises substantive questions about how Claude Code's internal resource allocation functions, specifically whether the platform dynamically escalates computational resources beyond a user's selected tier when task complexity demands it. The accompanying screenshot served as empirical evidence of the token and agent count, lending credibility to what would otherwise appear to be an extraordinary claim.

Claude Code, Anthropic's agentic coding and research tool, is designed to orchestrate multiple parallel subagents to complete complex, multi-step tasks more efficiently. This architecture — sometimes referred to as a multi-agent framework — allows the primary agent to delegate subtasks to specialized child agents simultaneously, dramatically accelerating throughput on research-intensive or computationally demanding work. The 100-agent figure and 2M+ token usage reported here are consistent with aggressive parallelization of a substantive research task, and represent one of the more dramatic public demonstrations of Claude Code's agentic capacity observed in user communities.

The central ambiguity in the user's post concerns the distinction between Anthropic's subscription tiers and the underlying computational behavior of the model. Anthropic offers Claude at multiple pricing tiers, including a "Max" plan and a higher "Ultra" designation, which differ primarily in usage limits and access to more intensive processing modes. The user's confusion is understandable: if a Max-tier session is consuming resources at a scale typically associated with Ultra, it raises questions about whether tier labels accurately reflect actual resource consumption or whether the system can silently exceed tier boundaries. Anthropic has not publicly documented automatic tier escalation behavior, making this a genuine gap in user-facing transparency.

The episode connects to a broader pattern in the AI industry wherein agentic systems, once deployed on open-ended tasks, can generate resource consumption that surprises even experienced users. As AI labs push toward increasingly autonomous, long-horizon task completion, the token economics of multi-agent orchestration become a critical consideration — both for pricing model design and for user cost management. The 2M+ token figure in under 20 minutes suggests that, at scale, agentic Claude Code sessions could impose significant API costs or exhaust subscription allocations far faster than users accustomed to single-turn interactions might anticipate.

This incident also underscores the growing importance of observability and user controls in agentic AI systems. Users operating Claude Code at high effort levels have limited real-time visibility into how many subagents are running, what resources are being consumed, or when usage thresholds are being approached — a transparency deficit that becomes increasingly consequential as these systems are applied to more complex and resource-intensive workloads. Anthropic and competitors in the agentic AI space will likely face mounting pressure to provide clearer dashboards, resource caps, and explicit user consent mechanisms as multi-agent deployments become routine rather than exceptional.