Proposal: Evidence Bench evidence-gated computational-science A2A sample

[kstawiski]

Proposal

I would like to contribute Evidence Bench as a computational-science A2A ecosystem sample:

• source: https://github.com/kstawiski/evidence-gated-scientific-agent
• first public release: https://github.com/kstawiski/evidence-gated-scientific-agent/releases/tag/v0.3.0
• license: Apache-2.0
• SDK/protocol: official a2a-sdk[fastapi] 1.1.0, Agent Card advertising A2A 1.0 JSON-RPC

Evidence Bench demonstrates an evidence-gated scientific agent rather than a generic chat agent. Qwen performs planning, retrieval, and sandboxed Python/R analysis; Gemma independently audits methods and results; deterministic code validates tool policy, artifacts, claims, and provenance. The service supports isolated persistent workspaces, raw A2A file parts, bearer-authenticated task execution, MCP retrieval tools, immutable per-workspace PyPI/CRAN/Bioconductor environments, and report/provenance artifacts.

Interoperability evidence

• repository A2A API tests pass in public CI;
• a deployed authenticated SendMessage task passed all 9 retrieval-grounding checks, including official SciPy/R sources, observed-URL enforcement, deterministic validation, and independent review;
• the Agent Card and JSON-RPC semantics are documented at https://github.com/kstawiski/evidence-gated-scientific-agent/blob/v0.3.0/docs/WEB_AND_A2A.md;
• the proposed contribution dossier is at https://github.com/kstawiski/evidence-gated-scientific-agent/blob/v0.3.0/docs/A2A_ECOSYSTEM_SUBMISSION.md.

Before preparing a significant PR, could maintainers advise which shape is preferred?

1. a small in-tree integration/sample that delegates to a running Evidence Bench instance; or
2. an external ecosystem/example entry linking to the full repository.

I am happy to reduce the contribution to the smallest useful A2A-specific example and follow the repository's test and documentation conventions.

[msampathkumar]

Hi @kstawiski Thank you for the interest.

I am currently working to revamp the code samples. For inspiration, you can check the helloworld example we have recently updated. After that you create a simple draft PR with pytest, then it would be great!

Tip: I would recommend to sticking to A2A v1.0 stable version.

AgentInterface(
            protocol_binding='JSONRPC',
            url='http://127.0.0.1:9999',
            protocol_version='1.0',         # Do not forget to add this line
        )

https://a2a-protocol.org/latest/tutorials/python/3-agent-skills-and-card/#agent-card

Good luck to you.