How Many Tools Can an AI Agent Handle at Once

The Direct Approach: All Tools, Every Call

When you pass all tool definitions to the model in every API call, the effective limit is around 10 to 15 tools. This is not a hard model limitation; it is a practical quality threshold. Below this number, the model can distinguish between tools reliably, the token overhead is modest, and selection accuracy stays above 95%. Above this number, three things degrade: selection accuracy drops as similar tools create ambiguity, token costs increase as definitions consume more of the context window, and response quality can suffer as the model's attention is split across a large input.

The exact number varies by model, schema quality, and the degree of overlap between tools. An agent with 20 highly distinct tools (no two are similar in purpose) may perform as well as an agent with 10 overlapping tools. Well-written descriptions that clearly disambiguate similar tools push the practical limit higher. Vague descriptions that leave the model guessing push it lower.

The token cost of the direct approach scales linearly with tool count. Each tool definition consumes 100 to 500 tokens depending on schema complexity and description length. With 10 tools averaging 250 tokens each, you add 2,500 tokens to every API call. With 50 tools, that jumps to 12,500 tokens, which is a significant fraction of the context window and a meaningful cost multiplier at scale. At 100,000 conversations per month with Claude Sonnet, the difference between 10 tools and 50 tools is roughly $3,000 per month in tool definition overhead alone, before accounting for any tool execution costs.

What Happens When You Exceed the Threshold

The failure mode when you pass too many tools is not a hard error. The model does not crash or refuse to respond. Instead, quality degrades gradually. The first symptom is usually increased tool selection errors: the model picks a plausible but wrong tool, especially when two tools serve similar purposes. A search_customers and a find_customer_by_email look similar enough that the model may pick the wrong one when the user's intent is ambiguous.

The second symptom is unnecessary tool calls. With many tools in context, the model sometimes calls a tool when it could have answered directly from conversation history or its own knowledge. The tool definitions occupy attention that would otherwise focus on the user's message, leading the model to over-index on tool use as a problem-solving strategy.

The third symptom is argument quality degradation. When the model is reasoning about which of 40 tools to use, it has less reasoning capacity left for constructing accurate arguments. Parameter hallucination (inventing values that were not in the conversation) becomes more common because the model's attention is spread thinner across a larger input.

With Routing: Hundreds of Tools

Dynamic tool routing removes the practical limit by selecting a subset of relevant tools for each query. An agent with 200 registered tools might present only 8 to the model for any given interaction, keeping token costs low and selection accuracy high. The routing layer analyzes the user's message, determines which tools are most likely to be relevant, and passes only those tools to the model.

This is how production systems at scale operate. Enterprise AI assistants often have dozens or hundreds of tool integrations spanning CRM, project management, document management, analytics, and communication systems. Routing ensures the model sees only the tools relevant to the current conversation, regardless of the total tool count.

Routing strategies range from simple to sophisticated. Keyword matching maps words in the user's message to tool categories and costs almost nothing in latency. Embedding-based similarity compares the user's query against tool descriptions in vector space and adds 50 to 100ms. Intent classification uses a lightweight model to determine the user's goal and map it to a tool set, adding 100 to 200ms. Combined approaches use multiple signals with weighted scoring. The tool router guide covers implementation details for each strategy.

Model Differences

Different models handle large tool sets with varying effectiveness. Claude's models perform well with up to about 15 to 20 tools in the direct approach and have strong parallel tool use support for fan-out patterns. GPT-4's strict mode can help with larger tool sets by enforcing exact schema compliance, reducing parameter errors even when the model's attention is split across many definitions. Gemini handles tool definitions somewhat differently due to its function declaration format, and benchmarks show similar practical limits in the 10 to 15 range for direct use.

All models benefit more from better schemas than from model-specific optimizations. A set of 12 well-designed tools with clear, distinct descriptions will outperform a set of 8 poorly designed tools on any model. If your tool selection accuracy is low, improving your schemas is almost always more effective than reducing your tool count.

Memory Reduces the Effective Tool Count

Persistent memory helps with tool scaling in a less obvious way: by remembering which tools a user typically needs. If memory shows that 80% of a particular user's interactions require the same 5 tools, the routing layer can pre-select those tools with high confidence, reducing the selection problem to a small, familiar set. Adaptive Recall's cognitive scoring naturally supports this because frequently used tool memories receive higher activation scores, creating a user-specific "favorite tools" list that improves routing accuracy for repeat interactions.

Memory also helps with tool selection disambiguation. When the routing layer presents 8 tools and two of them overlap in purpose, past outcome memories about which tool the user has successfully used before act as a tiebreaker. The model recalls "For this user, the detailed_order tool produced better results than the order_summary tool last time" and makes the correct selection without needing to reason from the schema alone.