SPEAKER_1: Alright, so last lecture we established that tools are what make an agent real — the schema is the interface between reasoning and the world. Today I want to pull on something that's been building across this whole course: what happens when agents from completely different companies, different stacks, need to work together? SPEAKER_2: That's exactly where interoperability becomes the central architectural challenge. And it's not theoretical anymore — interoperability in collaborative agent infrastructure means AI agents from different systems can communicate, share data, and coordinate actions seamlessly across platforms. The question is how you make that trustworthy. SPEAKER_1: So what's the actual state of this right now? How many agentic systems are even capable of cross-infrastructure collaboration today? SPEAKER_2: Fewer than most people assume. The honest answer is that true cross-infrastructure interoperability is still maturing — most enterprise deployments are siloed within a single vendor's stack. But the infrastructure to change that is arriving fast. MCP now has over ten thousand active servers globally and ninety-seven million monthly SDK downloads, with OpenAI, Google DeepMind, Microsoft, and AWS all supporting it. That's the foundation layer. SPEAKER_1: Right, we covered MCP back in lecture two. However, connecting agents across company boundaries presents unique challenges. What makes the cross-boundary case so much harder? SPEAKER_2: Trust is the core problem. Inside one organization, you control the identity layer — you know which agent is which, what permissions it has, what it's allowed to touch. The moment an agent crosses an organizational boundary, none of that is guaranteed. You're essentially asking: how does Agent A at Company X verify that Agent B at Company Y is who it claims to be, and that it hasn't been compromised? SPEAKER_1: So how do we ensure identity verification for agents across different infrastructures? SPEAKER_2: Not at the agent layer, no. API keys authenticate a service, not an agent's identity, capabilities, or current permission scope. That's why NIST launched its AI Agent Standards Initiative in February 2026 — specifically to address interoperable and secure AI agents through industry-led standards. They're building on three pillars: industry-led standards, open-source protocols, and dedicated AI agent security research. They're even holding sector-specific listening sessions starting April 2026. SPEAKER_1: The Linux Foundation's involvement with the Agentic AI Foundation highlights the importance of neutral governance in interoperability. SPEAKER_2: Demand for neutral governance. When MCP was donated to the Linux Foundation in December 2025, it signaled that no single vendor should own the interoperability layer. Ninety-seven new members in one month reflects organizations recognizing that if they don't help shape the standard, they'll be subject to one they didn't design. It's the same dynamic that drove open-source database standards a decade ago. SPEAKER_1: In cross-infrastructure scenarios, how do MCP, A2A, and ACP facilitate agent collaboration? SPEAKER_2: Clean separation of concerns. MCP handles agent-to-tool connections with strong audit trails — it excels in single-agent tool access. A2A standardizes multi-agent coordination across organizational boundaries. ACP handles lightweight messaging for scalable enterprise collaboration and suits legacy systems with REST patterns. The hybrid approach combines MCP for tool connections and A2A for agent coordination — that's what most serious enterprise deployments are converging on. SPEAKER_1: IBM's watsonx Orchestrate exemplifies governed interoperability across platforms. SPEAKER_2: It's a concrete implementation of governed interoperability. Watsonx Orchestrate enables agents to connect across SAP, Salesforce, ServiceNow, and custom systems under a single governance layer. The key mechanism is agent catalogs — they define approved agents, their permissions, and their integrations. That catalog becomes the backbone for orchestrating reusable agents across departments without every team rebuilding trust from scratch. SPEAKER_1: That catalog idea is interesting — it's almost like a verified directory. How does that connect to what HiClaw is doing with AgentScope? SPEAKER_2: HiClaw is a real-world example of the Manager-Workers architecture applied to cross-infrastructure collaboration. They joined AgentScope to partner with CoPaw specifically for building multi-agent infrastructure that spans different intelligence cores — CoPaw, OpenClaw, ZeroClaw. The key insight is that HiClaw allows creating Managers and Workers using diverse underlying models while maintaining consistent agent behavior. It's interoperability at the agent-design layer, not just the protocol layer. SPEAKER_1: And Microsoft open-sourced an Evals kit for agent interoperability in February 2026 — why does evaluation matter so much here specifically? SPEAKER_2: Because interoperability claims are easy to make and hard to verify. Microsoft's Evals for Agent Interop starter kit tests agents in realistic work scenarios — not synthetic benchmarks. When agents cross infrastructure boundaries, subtle failures emerge: mismatched schemas, permission scope mismatches, latency-induced state drift. You need evaluation tooling that surfaces those failures before they hit production. Open-sourcing it accelerates enterprise adoption by giving teams a shared baseline. SPEAKER_1: What about payments? If an agent from one company is genuinely hiring an agent from another to complete a subtask, how does that transaction actually work? SPEAKER_2: This is where the agentic economy gets concrete. Cross-boundary payments for agentic services are still early, but the pattern emerging is micropayment rails tied to task completion — similar to what we covered with SWM tokens in lecture nine. An orchestrating agent authorizes a budget, the external agent completes the task, and settlement happens programmatically. The governance challenge is ensuring the authorizing agent actually had permission to spend that budget in the first place. SPEAKER_1: So the scalability question — does interoperability actually make systems more scalable, or does the overhead of cross-boundary trust eat those gains? SPEAKER_2: It multiplies value when done right. Interoperability means an agent doesn't have to be rebuilt for every new data source or department — it connects. IBM's framing is precise: interoperability multiplies AI agent value by enabling collaboration across workflows, data sources, and departments. The overhead is real, but it's a one-time investment in the trust and protocol layer rather than a recurring cost per integration. SPEAKER_1: So for Suri and everyone working through this course — what's the architectural truth they should carry forward from this? SPEAKER_2: The future of AI isn't one company's agents doing everything — it's agentic bridges, where an agent in one company's infrastructure can securely hire, verify, and collaborate with an agent from another. That requires standardized protocols, verified identity, governed catalogs, and neutral standards bodies. The organizations investing in that interoperability layer now are the ones whose agent infrastructure will compound in value. Everyone else will be rebuilding integrations indefinitely.