Multi-Agent AI, Open Protocols & Startup Acceleration with Saurabh Tiwary | AI Basics

Episode Summary

Executive Summary: The episode centers on the shift from chat-based LLM use to autonomous AI agents that can perform multi-step, ambiguous work. Saurabh Tiwari explains Google Cloud’s ADK and A2A protocol as open infrastructure for building interoperable agents, emphasizes quality, trust, permissions, and human-in-the-loop controls, and highlights multimodal agents as a major next step for enterprise automation and customer support.

Main Topics: From chatbots to agents (Priority: 5/5): The conversation frames the key paradigm shift: moving beyond prompt-and-response LLM use toward agents that can decide next actions, handle multi-step work, and act on ambiguous tasks. Trust, quality, and production readiness (Priority: 5/5): Tiwari argues that agent adoption depends on reliability—especially task completion rates and accuracy—before companies will delegate important work to them. Google’s ADK for agent building (Priority: 5/5): The Agent Development Kit (ADK) is presented as an open-source framework for creating agents with tracing, debuggability, latency controls, and model/cloud flexibility. A2A: agent-to-agent interoperability (Priority: 5/5): The Agent-to-Agent protocol is described as an open standard for allowing agents built on different platforms to communicate across organizations and systems. Human-in-the-loop and permissions (Priority: 4/5): The discussion stresses that agents need oversight, interrupt points, and permission boundaries for sensitive business actions like payroll, invoices, and account updates. Multimodal AI in workflows (Priority: 4/5): Tiwari highlights multimodality—combining text, speech, image, and video—as a major capability that will make agents more useful for tasks like receipts, storefront setup, coaching, and support. Enterprise examples and ecosystem adoption (Priority: 4/5): Examples from internal use cases and customers like Shopify illustrate how agents are already being used in production-like workflows and how ecosystem standards are gaining traction.

Key Arguments: Agents are evolving from simple prompt wrappers into systems that can take actions, evaluate outputs, and determine the next steps in open-ended workflows. Real-world delegation requires high quality; agents with around 50% accuracy are not trustworthy for important work, while 80-90% accuracy starts to make them viable. Open-source infrastructure matters: ADK lets developers build agents with tracing, debugging, and deployment flexibility across models and clouds. Interoperability is essential because the ecosystem will not converge on one framework, so A2A enables different agents to communicate regardless of how they were built. Human-in-the-loop controls remain necessary for sensitive, high-stakes actions, especially where security, permissions, and financial or payroll changes are involved. Multimodal understanding will expand agent usefulness by letting systems reason over screens, receipts, calls, video, and other non-text inputs. Agent adoption will happen incrementally, with simpler chores and normalization tasks being automated first before more complex delegated actions. The industry is already seeing production deployments and ecosystem support from major vendors, suggesting standards are coalescing quickly.

Data Points: AI voices referenced in Google report: 23 - The host promotes Google Cloud’s report, which includes insights from 23 leading voices in AI. Customer support / data normalization time saved: 15 to 30 minutes - The founders-update example suggests AI can save analysts time by normalizing updates and identifying missing data. Accuracy threshold for interest: 80-90% - Tiwari says agents become interesting and trustworthy once they reach this range of accuracy. Low-trust accuracy example: 50% - He notes that an agent correct only one out of two times is not viable for important tasks. A2A partners at launch: 50 to 60 partners - When A2A was announced at Cloud Next, roughly 50-60 partners supported the protocol. A2A partners after growth: 120 partners - Tiwari says support for the protocol grew to about 120 partners. A2A governance transfer: Linux Foundation - The protocol was donated to the Linux Foundation for open governance. A2A announcement timing: April of this year / month two - He says the protocol was announced at Cloud Next in April and is still very early in its lifecycle.

Pivotal Quotes: "we are basically asking these agents to make indeterminate tasks" — Saurabh Tiwari: Explaining the new definition of agents and how they differ from simple prompt wrappers or workflows. "if you are executing or calling that agent to do certain tasks, how often does it finish" — Saurabh Tiwari: On why quality and completion rate are the foundation of trust in production agents. "Clerical work, chores, they're going to go away" — Saurabh Tiwari: Describing the expected business impact of agent-to-agent automation across mundane operational tasks.

Implications: AI agents are moving from novelty to infrastructure. Expect near-term gains in workflow automation, customer support, and operations, but only where quality, permissions, interoperability, and human oversight are built in.