Signal Snapshot

Teams are starting to combine open runtimes and managed platforms inside one architecture

The agent stack is already moving beyond single-vendor product comparison. OpenAI publishes core building blocks for agents, AWS shows both multi-agent collaboration and human-in-the-loop patterns, Google pairs multi-system agents with A2A, and Microsoft pushes Semantic Kernel Agents to GA. Anthropic's Code with Claude page is an event announcement rather than an enterprise runtime launch, but it still reinforces how much attention tool-using developer agents are attracting.

11

Published evidence

Only papers and official announcements / docs that directly support the article's claims are listed.

20+

Research pool

Candidate URLs were limited to primary sources available by publication.

3 layers

Converging design questions

The real design work was separating protocol, hosted runtime, and approval paths.

What Stood Out

The strongest signals

Google pushed protocol and platform at the same time

Vertex AI multi-system agents and A2A together suggested that managed platforms and open coordination rules should be designed as complementary layers. The important shift was that agents were no longer framed as something that stays inside one cloud boundary forever.

AWS and Microsoft made managed orchestration more concrete

AWS paired multi-agent collaboration GA with explicit human approval patterns. Microsoft made a similar move by treating role-based agent orchestration as a stable SDK surface in Semantic Kernel Agents GA rather than a purely experimental concept.

OpenAI and Anthropic raised expectations around developer agent loops

OpenAI's agent-building releases and Operator made web, tool, and computer-use loops far more visible. Anthropic's Code with Claude announcement was not a full enterprise runtime release, but it still showed how quickly tool-using coding workflows were moving into the center of attention.

Use Cases

Use cases that look practical

Browser-assisted research and application prep

  • Browser-oriented agents made it easier to gather information and prepare form-based work.
  • The safer near-term pattern was still automation plus confirmation, not fully unsupervised submission.

Internal assistants that span multiple systems

  • Search, FAQ, data retrieval, notifications, and approval requests could be split across specialized workers.
  • Using orchestration SDKs and open coordination rules made later integration changes easier to absorb.

Concrete Scenarios

Specific scenarios already visible in the source set

AWS used PTO approval flows to define the boundary between automation and approval

The AWS human-in-the-loop article showed an HR agent that asks for user confirmation before creating, updating, or canceling time-off requests, and a return-of-control pattern for cases that need richer user edits. The practical message was that useful agents are not the ones that do everything automatically, but the ones that can be stopped at the right moment.

A2A and multi-system agents made specialist coordination easier to picture

Google's A2A announcement used a candidate-sourcing example where a client agent delegates tasks to specialist remote agents and follows a long-running task lifecycle. Read alongside AWS supervisor patterns and AutoGen v0.4, this made research, retrieval, validation, and synthesis look like separable agent roles rather than one monolithic loop.

Operating Implications

What teams needed to decide early

Observation

The harder question is no longer which vendor to pick, but where to separate protocol, hosted execution, and approvals.

  • Define tool schemas and message contracts so runtime choices can change without breaking everything.
  • Separate read-only automation from write actions, and put confirmation or return-of-control in front of changes.
  • Distinguish the steps that fit a single agent from the ones that justify a supervisor-plus-specialist pattern.
  • Do not read event announcements and GA releases as if they carry the same maturity signal.

Key Takeaway

Conclusion

The agent stack is becoming an architecture problem: how to combine open coordination rules, managed execution, and human approval, rather than how to choose one product.