Agent Sense — Episode 8: Governed Catalog of AI Assets

Enterprise AI agents does not fail from lack of ideas. They fail from agent sprawl: duplicate builds, unclear ownership, no way to know what is safe to reuse. Jyotsna Narayanan, Principal Product Manager for watsonx Orchestrate at IBM, joins host Monika Aggarwal to break down what a governed catalog of AI agents, tools, and MCP servers needs to include, how to evaluate reuse readiness using references like OWASP Top 10 and CoSAI, and where to start.

Topics: AI agents, agentic AI, enterprise AI governance, AgentOps, MCP, Model Context Protocol, watsonx Orchestrate, agent catalog, AI reuse, enterprise automation, production AI.

Vendor-agnostic, not monetized, based on field experience. Views are personal and do not represent IBM.

The episode is about hashtag#9minutes long. In this episode, my co host Frank Chávez and I are joined by Vitalii Duk from Dynamiq to discuss how agents are moving from answering questions to doing work across tools, systems, and runtime environments.

“𝑪𝒉𝒂𝒕 𝒎𝒆𝒂𝒏𝒔 𝒂𝒏 𝒂𝒈𝒆𝒏𝒕 𝒄𝒂𝒏 𝒂𝒏𝒔𝒘𝒆𝒓 𝒚𝒐𝒖𝒓 𝒒𝒖𝒆𝒔𝒕𝒊𝒐𝒏. 𝑪𝒐𝒎𝒑𝒖𝒕𝒆𝒓 𝒖𝒔𝒆 𝒎𝒆𝒂𝒏𝒔 𝒕𝒉𝒆 𝒂𝒈𝒆𝒏𝒕 𝒄𝒂𝒏 𝒕𝒂𝒌𝒆 𝒂𝒄𝒕𝒊𝒐𝒏 𝒂𝒄𝒓𝒐𝒔𝒔 𝒂𝒑𝒑𝒍𝒊𝒄𝒂𝒕𝒊𝒐𝒏𝒔, 𝒕𝒐𝒐𝒍𝒔, 𝒂𝒏𝒅 𝒘𝒐𝒓𝒌𝒇𝒍𝒐𝒘𝒔. 𝑻𝒉𝒂𝒕 𝒊𝒔 𝒘𝒉𝒆𝒓𝒆 𝒆𝒏𝒕𝒆𝒓𝒑𝒓𝒊𝒔𝒆 𝑨𝑰 𝒔𝒕𝒂𝒓𝒕𝒔 𝒕𝒐 𝒃𝒆𝒄𝒐𝒎𝒆 𝒐𝒑𝒆𝒓𝒂𝒕𝒊𝒐𝒏𝒂𝒍.”

We share a practical frame:
🔹 Start with bounded, observable, and measurable agent use cases
🔹 Define the agent’s authority boundary before giving it access to systems
🔹 Use orchestration to manage work across tools, agents, approvals, and exceptions
🔹 Add observability, audit, and human review before scaling agent action
🔹 Treat Day 2 and Day 3 readiness as part of the design, not an afterthought

𝙏𝙝𝙞𝙨 𝙞𝙨 𝙩𝙝𝙚 𝙚𝙣𝙩𝙚𝙧𝙥𝙧𝙞𝙨𝙚 𝙖𝙧𝙘𝙝𝙞𝙩𝙚𝙘𝙩𝙪𝙧𝙚 𝙨𝙝𝙞𝙛𝙩: 𝙖𝙜𝙚𝙣𝙩𝙨 𝙖𝙧𝙚 𝙢𝙤𝙫𝙞𝙣𝙜 𝙛𝙧𝙤𝙢 𝙧𝙚𝙨𝙥𝙤𝙣𝙨𝙚 𝙩𝙤 𝙚𝙭𝙚𝙘𝙪𝙩𝙞𝙤𝙣. 𝙏𝙝𝙚 𝙫𝙖𝙡𝙪𝙚 𝙞𝙨 𝙣𝙤𝙩 𝙤𝙣𝙡𝙮 𝙞𝙣 𝙬𝙝𝙖𝙩 𝙩𝙝𝙚 𝙖𝙜𝙚𝙣𝙩 𝙘𝙖𝙣 𝙙𝙤. 𝙄𝙩 𝙞𝙨 𝙞𝙣 𝙝𝙤𝙬 𝙨𝙖𝙛𝙚𝙡𝙮, 𝙤𝙗𝙨𝙚𝙧𝙫𝙖𝙗𝙡𝙮, 𝙖𝙣𝙙 𝙧𝙚𝙥𝙚𝙖𝙩𝙖𝙗𝙡𝙮 𝙞𝙩 𝙘𝙖𝙣 𝙤𝙥𝙚𝙧𝙖𝙩𝙚 𝙞𝙣 𝙧𝙚𝙖𝙡 𝙗𝙪𝙨𝙞𝙣𝙚𝙨𝙨 𝙬𝙤𝙧𝙠.

Episode 6 of Agent Sense continues from Episode 5, where we talked about MCP, A2A, and enterprise integration.

In this episode, Monika Aggarwal and Frank Chávez discuss why enterprise agents need governed access to core systems before they can scale in production.

MCP helps agents connect to tools and systems. But connection alone is not enough. As agents start working across ServiceNow, Workday, SAP, HR, IT, finance, and customer operations, enterprises need a control layer.

That is where the MCP Gateway comes in.

We discuss how an MCP Gateway helps manage identity, policy, approvals, audit, and traceability. It gives agents access to approved tools without opening direct, unmanaged paths into core enterprise systems.

In about 4 minutes, we cover:

🔹 Why direct agent access to core systems creates risk
🔹 How MCP Gateway supports controlled enterprise access
🔹 Why public MCP servers are useful for testing, but not enough for production
🔹 How approved tools help agents scale across business workflows
🔹 Why traceability matters when agents take action

Episode 5 was about connection.
Episode 6 is about controlled access.

Disclaimer: The views shared are based on our personal experience and do not represent the views of IBM.

Tags for Spotify search:

Agentic AI, Enterprise AI, MCP, MCP Gateway, AI agents, ServiceNow, Workday, SAP, AI governance, agent governance, operational AI, enterprise architecture, AI integration, agentic workflows.

Theme: As AI systems evolve from single models into networks of autonomous agents, integration is the primary bottleneck.

Core Integration Challenges:

1. Fragmented Tool Access

2. Context Loss Across Agents

3. Tight Coupling & Low Reusability

4. Lack of Standardized Communication
   

   Integration needs standards.
   MCP standardizes how agents connect to systems.
   Agent to agent communication standardizes how they pass work.

   
   

   🔷 MCP connects agents to enterprise systems.
   🔷 A2A connects agents to each other so work can move across the enterprise.

Theme: Do autonomous databases fix bad data, or do they mainly improve operational reliability? Why are organizations moving toward autonomous operations?

In episode 3, we talked about an IT service agent that created operational noise during an outage. The AI agent acted fast, but the ownership and escalation data were wrong, so the actions were wrong.

If the data underneath these systems is fragile, should the data layer become autonomous too? In eposide 4 we are talking about autonomous databases, and what they can and cannot do in incidents like this.

I am Monika Aggarwal, AI Technical Practitioner. I build agentic workflows grounded in clear rules, good data, and governance. I am joined by my colleague Frank Chavez. He is a Technical Architect and hands-on builder specializing in multi-agent orchestration and AI integration patterns.

I bring the enterprise and operational view. Frank brings the engineering view.

Theme: Foundations & Data Readiness. Why do agents go rogue when the information source is weak?

This is episode three: Why IT Service Agents Fail in Production, A Data Readiness Problem. Most enterprise agentic failures are not related to the model. They are data failures. We are using a real IT service ticketing example to show why data readiness matters for agents.

I am Monika Aggarwal, AI Technical Practitioner. I build agentic workflows grounded in clear rules, good data, and governance. I am joined by my colleague Frank Chavez. He is a Technical Architect and hands-on builder specializing in multi-agent orchestration and AI integration patterns.

I bring the enterprise and operational view. Frank brings the engineering view.

This episode explores the line between deterministic business logic and autonomous agents in real business operations, using a Commercial and Investment Banking onboarding scenario to show where rules work, where agents help, and where humans must stay in control.

I am Monika Aggarwal, AI Technical Practitioner. I build agentic workflows grounded in clear rules, good data, and governance. I am joined by my colleague Frank Chavez. He is a Technical Architect and hands-on builder specializing in multi-agent orchestration and AI integration patterns.

I bring the enterprise and operational view. Frank brings the engineering view.

Enterprises often stop at building a clever prototype. But when that agent touches real systems, the question changes from Can it run? to Can we trust it?
That is the gap between experimentation and production.