AWS Strands Agents: How Tool Design Cut Tokens 96%

An AWS developer advocate reduced agent token consumption from 52,000 to 2,000 — a 96% reduction — without changing the underlying model or the task the agent needed to perform.

Context

Token efficiency is the quiet cost driver in agentic AI deployments. An agent that reasons poorly about which tool to call burns tokens on failed invocations and redundant context loading; an agent whose tool definitions are misaligned with task intent burns tokens before it even begins executing. AWS Strands Agents, the open-source SDK that Amazon launched earlier in 2026, provides the framework that Morgan Willis used to demonstrate how dramatically tool design choices affect token consumption — independent of model selection or prompt engineering.

The Detail

Willis began with a baseline implementation that mapped each API endpoint directly to an individual agent tool. A task as simple as looking up a customer’s latest invoice required five chained API calls, each surfaced as a separate tool the agent needed to reason about and invoke in sequence. At this baseline, the agent consumed approximately 52,000 tokens per task.

The first optimization replaced endpoint-per-tool design with intent-based tool design. Rather than exposing the individual API operations, Willis consolidated multiple related endpoints into single tools organized around task outcomes. The invoice lookup task collapsed from five sequential tool calls to one. Token consumption dropped to approximately 2,000 — a 96% reduction from the baseline. Willis’s framing captures the architectural principle: “Your agent is going to have a better time reasoning around what tool to use and when, because these tools are more aligned to a task and less aligned to data.”

The second iteration added semantic filtering via a Model Context Protocol gateway. Rather than presenting all 16 available tools to the agent on every query, the MCP layer filters to contextually relevant tools per request. This reduced token consumption by roughly half again, to approximately 1,000 tokens. The mechanism matters beyond cost: an agent reasoning over 16 undifferentiated tools is more likely to select incorrectly than an agent reasoning over the 3-4 tools that semantic context identifies as relevant to the current query.

The design principle Willis derives from both optimizations is scope specificity: narrowly scoped agents with intent-aligned tools outperform general-purpose agents given broad tool access on both efficiency and accuracy. This mirrors patterns from AWS’s earlier work on MCP tool selection, where reducing tool surface area improved both reasoning speed and reliability.

Source: The New Stack

Content created with AI assistance and reviewed for accuracy.