Artificial intelligenceOctober 12, 20263 min · updated October 15, 2026
Postman explains Agent Mode: lessons for connecting an agent to business APIs
Daillac research
AWS publicationD
On October 9, 2026, AWS and Postman published feedback on the architecture of Agent Mode with Amazon Bedrock. The text notably describes the selection of accessible tools and preparing the context. For a business integration, these choices influence the information the agent receives and the operations it can request.
Feedback on architecture, with a scope to respect
The AWS publication examines the use of Agent Mode in Postman. The 40 million developers mentioned refers to the Postman community. This number does not measure active users of Agent Mode and should not be used to announce its adoption in companies.
The feedback highlights the importance of limiting the tools provided to the useful context and maintaining checks for operations that change a state. It can serve as a starting point for an integration review, subject to the needs and tools specific to your application.
Provide the tools necessary for the task
Let's take a fictional example of searching for an order. The agent needs a search tool, a reading tool, and limited access to delivery information. Account deletion or price modification functions do not help complete this task. They can remain outside the exposed scope.
For each tool, define the settings and the expected outcome. Searching for an order should distinguish "no match," "multiple matches," and "service unavailable." If all these situations return an empty list, the model may interpret an outage as an absence of orders.
The MCP vs API for connected agents guide describes this approach to operations. The choice of protocol must preserve understandable business functions and user-friendly rights.
Keep a verifiable context
Associate the returned information with its source and date when these elements change the decision. An address found in an old document must not automatically replace the current CRM address. The agent can present the conflict for verification.
The context must also specify the task boundaries. A read request may lead to a correction suggestion, but the correction must go through its approval process. After a refusal, check that no write tool was called.
For testing, measure the quality of responses and operations performed. Add a tool with an error and an attempt to perform an out-of-scope action. The test may reveal a correct response accompanied by an incorrect action, or the opposite.
The of the first workflow with n8n guide provides test cases that can be used for a pilot. Our AI integration service can scope the tools and controls before opening access to your business applications.
Review the architecture of your AI integrations
Present your needs and the relevant tools to scope the useful checks for your project.
Sources & method
Official AWS and Postman publication of 2026-10-09, accessed October 12, 2026. Summary of documented facts and Daillac analysis; examples explicitly fictional, without provider test or assigned client result.
News feed
October 12, 2026
Development and hosting
Vercel reduces deployment history: what to check before October 23
October 12, 2026
Artificial intelligence
Claude Haiku 5.5: should you switch models for support and repetitive tasks?
October 12, 2026
Artificial intelligence