AWS and Postman detail how Agent Mode scales AI agents to 40 million developers on Amazon Bedrock
In short: AWS published a joint engineering write-up with Postman describing how Agent Mode, Postman's AI-native interface for API testing and documentation, was built to run in production at scale. Rather than model quality, the team found that tool sprawl and missing context were the biggest obstacles to reliable agent behavior. The post details architectural patterns including dynamic tool scoping, schema-based query access, dedicated context handlers, and how Amazon Bedrock provides model flexibility, cross-Region inference, data residency controls, and prompt caching.
This summary was generated automatically by AI from AWS Machine Learning's publication. It is our own text, not a copy of the original — facts, figures and quotes belong to the source, linked above and below.
What changed?
- 1Tool-selection errors rose once the visible toolset exceeded roughly 40 tools; the architecture now narrows over 170 tools to about 15 relevant ones per request using a vector database of tool embeddings
- 2Postman consolidated many narrow read tools into a single schema-aware query tool over ClickHouse tables, letting the agent generate its own complex queries instead of using dozens of single-purpose tools
- 3Context (user location, active entities, prior state) caused more failures than missing capabilities; Postman built dedicated context handlers per entity type instead of serializing raw interface data
- 4Agent Mode requires user approval before any action that modifies application state, and uses Amazon Bedrock Guardrails to redact PII before it reaches the model
- 5Amazon Bedrock supports multiple Anthropic Claude models for the same integration, letting Postman route fast models to high-volume tasks and larger models to complex reasoning without rebuilding the integration
- 6Cross-Region inference profiles (geographic or global) let Postman handle bursty traffic while keeping processing within a defined geography such as the US or EU when required
Why it matters
This is a rare detailed look at what actually breaks AI agents in production — not model quality, but tool catalog size and context design. For teams building agents with many function-calling tools, it offers concrete numbers (the ~40-tool threshold, 170→15 tool narrowing) and patterns worth replicating.
What it means for AI agents and contact centers
If you run voice AI agents with tool/function calling (CRM lookups, call transfers, scheduling, data queries), the finding that agents degrade past roughly 40 visible tools and that context quality matters more than tool count is directly applicable — consider dynamic tool scoping per call intent and purpose-built context summarization instead of passing full CRM/call records into the prompt. The use of Amazon Bedrock Guardrails for PII redaction before model calls is also relevant for contact-center compliance.
🧪 Worth Testing
The dynamic tool-scoping and context-handler patterns described could be tested against existing voice agent / contact-center tool stacks to see if they reduce latency and tool-selection errors when scaling beyond a handful of integrated tools.
Dynamic tool scoping and context handlers for AI agents· New
Sources
- AWS Machine LearningOfficialPrimary sourceOriginal article →„How Postman runs Agent Mode for 40 million developers on Amazon Bedrock“9 Oct 2026, 18:35
- Published by source
- 9 Oct 2026, 18:35
- Found by our system
- 9 Oct 2026, 18:39
- Summary generated
- 9 Oct 2026, 18:41
This article was written by AI from the original source. Facts, numbers and prices come from the source; missing values are marked “Not specified”. Legal notice, copyright and privacy