Agent Builder: A Self-Service, Governed Way to Build Agents
Snehil opened with an overview of Agent Flow and Agent Builder, emphasizing that Agent Builder is a self-service tool that lets teams build custom agents without engineering support, while still maintaining validation, governance workflows, security scans, and versioning. Agent Builder has been generally available since the 2026.1 release. The agent creation workflow starts with authors drafting agent specifications, testing against live data, and iterating on instructions until the agent meets requirements.
Three Levels of Agent Instructions
A central theme of the session was the progression of agent instructions through three maturity levels:
-
Minimal prompt: A basic prototype used to validate the core idea quickly
-
Specific prompt: A production-ready version with added guardrails, exception handling, and structured output formatting
-
Configurable prompt: A flexible version using parameterized values, ideal for supporting multiple use cases from a single agent
Snehil noted that specific prompts are recommended for production use, while configurable prompts offer greater reuse across teams and scenarios.
Live Build: Website URL Enrichment Agent
To bring these concepts to life, Snehil built a Website URL Enrichment Agent for organization entities from scratch. The demonstration covered:
-
Creating the agent and selecting relevant MCP tools and web search capabilities
-
Testing the agent against live data, including a real example enriching the website URL for JPMorgan Chase Foundation
-
Watching the agent evolve from a minimal prompt to a specific prompt, adding guardrails and confidence levels with source citations
-
Reviewing built-in security guardrails that flag or block inappropriate instructions before they reach production
The enhanced version of the agent successfully returned a well-formatted recommendation with confidence scoring and citations before updating the entity's attributes.
Publishing, Review, and Configurable Deployment
Snehil detailed the roles of authors and approvers in the publishing process, along with the automated security scanning that runs before any agent goes live. Approvers receive a full system audit and can test the agent's prompt themselves before granting approval.
He also introduced Agent Deployment Packs (ADPs), which package configurable parameters and values so a single agent can be deployed flexibly across different scenarios. Rounding out the session, Snehil demonstrated the AI-assisted Agent Blueprint feature, which generates an initial set of instructions based on a natural language description of the desired agent — helping teams move past the blank-page problem when getting started.
What This Means for Your AI Journey
The session reinforced a practical, repeatable path for teams at any stage of their agent-building journey: start with a minimal prototype, iterate toward a specific prompt for production reliability, and move to a configurable prompt when a single agent needs to support multiple use cases. Whether you're a data steward, business analyst, solution architect, or product manager, this progression offers a clear framework for turning an idea into a working, governed agent.
Please see the attached pdf document for further reference, and Snehil's GitHub link.
#CommunityWebinar
#Featured
#AIML