A2A Protocol, AgentCore Runtime isolation, StructuredOutputSimilarity evals

Nov 12, 2026•Items: 3

Systems & planning

  • A2A Protocol connects LangChain and Strands agents…

    TL;DR: A2A connects a LangChain enterprise orchestrator and a Strands SaaS agent across AWS accounts, with OAuth authentication and bidirectional collaboration.

    • The reference architecture runs three agents in two AWS accounts: a LangChain orchestrator and context agent on ECS with Fargate, and a Strands compliance agent on Bedrock AgentCore.
    • The orchestrator fetches the SaaS agent’s Agent Card at startup and creates LangChain tools from its published skills; it sends A2A JSON-RPC requests over HTTPS using raw httpx, without depending on the Strands SDK.
    • Each side uses its own Amazon Cognito user pool. The SaaS agent verifies the registered callback Agent Card, obtains a token from its own pool, and calls the enterprise context agent over A2A. Onboarding exchanges public values, not secrets.
    • The compliance demo uses synthetic findings. The deployment notebooks require two AWS accounts in the same supported Region and run in order; the enterprise notebook takes about 20 minutes.

    Why it matters: This shows a concrete interoperability path where an orchestrator can discover and invoke a partner agent via an Agent Card and JSON-RPC, without importing the partner’s SDK or sharing internal implementation details. Separate Cognito user pools plus callback registration keep cross-account calls authenticated while limiting onboarding to public configuration values.

    Signals: The orchestrator supports a single OAuth client today; adding a second ISV requires one client and scope per ISV, adding its Agent Card URL, and restarting the orchestrator.

    Source: https://aws.amazon.com/blogs/machine-learning/build-composable-interoperable-multi-agent-systems-with-a2a-on-aws/

  • AgentCore Runtime

    TL;DR: AgentCore runtime delivers hardware-isolated agent sessions, faster start times, data isolation, and usage-based billing.

    • AgentCore runtime uses hardware to isolate agent sessions, separating workloads at the session level.
    • Supports faster agent session startup compared to previous versions.
    • Implements session isolation that also extends to user data.
    • Introduces usage-based billing for agent workloads.

    Why it matters: Extending the trust boundary to mounted data lets teams govern session data alongside isolated compute.

    Signals: It remains uncertain how isolation guarantees hold under extreme concurrency or in the presence of advanced evasion techniques.

    Source: https://aws.amazon.com/blogs/machine-learning/run-agents-securely-and-efficiently-with-the-next-generation-agentcore-runtime/

Evaluation, grounding & uncertainty

  • StructuredOutputSimilarity provides field-level…

    TL;DR: StructuredOutputSimilarity scores structured agent outputs field by field using deterministic comparisons; its default scoring path requires no LLM judge.

    • Field-by-field comparison isolates partial correctness and handles variations such as reordered list items and numeric differences within tolerance.
    • Stickler uses the existing Pydantic model as the evaluation schema and infers comparators from field types and names, with optional overrides.

    Why it matters: For agents expected to emit typed objects, deterministic per-field scoring makes CI regressions reproducible without adding an LLM judge to the evaluation loop. The per-field breakdown also narrows debugging to specific fields (e.g., list ordering vs numeric tolerance), while semantic or LLM-based comparators can be enabled only when those additional dependencies and runtimes are acceptable.

    Source: https://strandsagents.com/blog/structured-output-similarity-strands-evals/

Subscribe

Get Applied Signals weekly newsletter in your inbox