Hello Apigee Community! 
In our latest TechTalk, Tom Hendrix and Brandon Verzuu from AppyThings teamed up with Kevin Bouwmeester from Google Cloud to explore how organizations are moving autonomous AI agents from experimental MVPs to scaled production systems.
The session focuses on how your integration layer acts as the critical control point for agent actions, detailing how to turn your existing Apigee architecture into a robust AI Gateway. 
Watch the recording
Access the presentation slides here
The agentic control layer 
To scale autonomous agents safely, enterprises must establish rigid boundaries between agent runtimes and core systems. Our speakers highlighted three core pillars of this architectural pattern:
-
Centralized model and tool governance: Using Apigee to sit in the middle of model queries and tool calls. This centralized placement enables intelligent model routing (optimizing for cost, quality, and latency), early policy checks via Model Armor, and semantic caching to minimize token costs.
-
Preventing the confused deputy vulnerability: Enforcing secure user-delegation profiles (OAuth 2.0) and introducing automated, ultra-short-lived automated identities (Spiffy/SPIRE sidecars) to govern high-speed east-west agent interactions.
-
Preserving experience consistency with MCP Apps: Moving beyond raw text extraction by utilizing the Model Context Protocol (MCP) Apps extension. This allows Apigee to bundle interactive, on-brand frontend components with backend API responses, ensuring external agents render structured interfaces instead of plain-text summaries.
Key takeaways from the session 
Our speakers provided an in-depth breakdown of practical enterprise patterns based on their recent pilot work with MTN (South Africa’s telecom group handling over 75 billion annual API calls):
-
API transcoding to MCP: Apigee makes it easy to transcode existing REST/SOAP endpoints into MCP-ready servers. This lets developers wrap traditional business logic into clean “tools” that agents can discover and call at runtime.
-
Curation via API products: To prevent “context blow”—where giving an agent too many tools leads to slow execution times and hallucinations—the team demonstrated how to bundle tools into specialized AI products in Apigee, exposing only the exact scopes an agent requires.
-
Isolating non-deterministic behavior: During the Q&A, the engineers addressed the challenges of testing agents. A key strategy is separating non-deterministic LLM predictions from predictable operations (like mathematical calculations or database lookups) by delegating those calculations back to traditional API tools.
Next steps 
-
Continue the conversation: How is your team addressing token cost optimization or identity delegation for agentic workflows? share your thoughts and questions in the comments below!
-
AI gateway capabilities: Access to high-quality tools, such as enterprise APIs, is essential for agents to be effective. In addition to traditional API management capabilities, Apigee provides AI gateway related features to govern and optimize your AI traffic while transforming existing services into discoverable tools for your agents
AI gateway capabilities | Apigee | Google Cloud Documentation -
Featured Resource: AppyThings AI Whitepaper & Assessment
We highly recommend downloading AppyThings’ new release, “Why AI Agents Stall Before They Reach Production”, to explore technical frameworks to scale your agentic pipelines. You are also invited to book a complimentary AI architecture assessment Why AI Agents Stall Before They Reach Production -
Connect with Google Cloud: Speak with a representative to explore how Google Cloud and AppyThings can jointly help scale your enterprise AI initiatives
Contact Us | Google Cloud -
Share your feedback: Help us refine future technical developer meetups by sharing your feedback here
Feedback: Apigee Community TechTalks -
Stay tuned for future community sessions
Google Cloud Apigee Community TechTalks