Todd Segal designs infrastructure for conversations that might not finish. As a Principal Engineer at Google Cloud, he co-led development of the A2A protocol—a standard for how autonomous agents coordinate when neither party can assume the other will stick around.

A2A's central abstraction is a stateful Task object that persists across whatever happens next: the delegating agent going offline, the receiving agent failing partway through, both systems disappearing and reappearing multiple times before work completes. An agent might delegate a task and hear nothing back for hours, days, or potentially never.

What Persistence Requires When Systems Vanish

In an April 2025 podcast, Segal explained why A2A needed fundamentally different assumptions than tool-calling protocols:

Real workflows involve human review, external dependencies, research that can't be rushed. The protocol handles this through push notification mechanisms—the client agent provides a secure callback URL, and the receiving agent notifies when work completes.

What happens when the callback URL is no longer valid because the delegating agent restarted? How do agents verify that a notification actually came from the system they delegated to? What does retry logic look like when you don't know if the other agent is still working or abandoned the task hours ago?

Segal made the Task object itself the source of truth. Agents can query task status, retrieve partial results, understand what happened even when the original connection is long gone. Disappearance becomes something the protocol expects and handles.

Standardizing Systems That Keep Changing

A2A builds on HTTP, Server-Sent Events, and JSON-RPC. That choice pushes certain problems to implementers: connection management, authentication schemes, rate limiting all happen at layers A2A doesn't specify. The flexibility matters when systems are still evolving.

The A2A GitHub repository shows what that produces. The roadmap includes dynamic UX negotiation—letting agents agree on information exchange formats at runtime. One agent might return structured JSON; another might generate natural language summaries. The protocol accommodates both.

Dynamic capability checking through QuerySkill() methods, authorization schemes that can evolve in AgentCard specifications, streaming improvements for partial results. Agents will develop new skills and need to advertise them without waiting for protocol updates.

What Coordination Looks Like at the Edge of Stability

By June 2025, more than 100 technology companies supported A2A under Linux Foundation governance. Organizations are encountering coordination failures at scale—problems that only become visible in production when agents delegate tasks across systems that have never interacted before.

Segal's work involves defining communication patterns while capabilities are still emerging. The protocol uses "agent cards" for capability advertisement combined with multiple registries for discovery—trust establishment for systems that have never interacted and might never interact again. Service discovery plus the infrastructure for agents to negotiate, delegate, wait, and recover from failures that might not surface until days after a task was delegated.

Agents vanish mid-task. Networks partition. Systems restart without warning. Segal designs coordination infrastructure that expects these conditions and persists through them. Reliability here means surviving failures, maintaining state across disappearances, recovering when participants return.