Scientific Agent Skills
Scientific Agent Skills 未来与实践项目：构建可复现研究 Copilot
以可移植证据包、评估治理和数据库扩展为主线，规划一个可复现的研究代理实践项目。
你将学会
- Design a provenance-aware capstone
- Define evaluation and governance gates
- Separate current evidence from future proposals
开始前需要
- Basic Git and command-line usage
- Comfort reading a project README
You can explain the project, run its documented first step, and decide what to verify before adopting it.
先看结论
- Future value comes from portable evidence bundles and explicit capability manifests.
- A narrow provenance-aware copilot is a useful capstone project.
- Evaluation must include citation accuracy, abstention, adversarial fixtures, cost, and human review.
The next useful abstraction
Scientific Agent Skills 的下一步不只是增加技能，而是让每个技能输出带版本、标识符、原始响应、引用、质量标记和人工决策的证据包；本文据此设计一个窄领域、可回放、可审计的研究 Copilot 项目，并明确哪些内容仍属于未来提案。
A shared manifest could declare capability, data classification, allowed domains, expected latency, and failure modes. Hosts would then discover not only what a skill can do, but also what it is allowed to touch.
Capstone project: a provenance-aware research copilot
Build a small service that accepts one research question, selects one pinned skill, runs a public-data fixture, and stores an evidence bundle. Add a reviewer UI that shows the query, identifiers, source links, database version, raw response hash, model draft, and approval state side by side.
Keep the first release intentionally narrow: one domain, one database adapter, one model, and one export format. The project is valuable when a second person can replay the fixture and explain every field, not when it claims broad scientific coverage.
Evaluation and governance roadmap
Measure citation completeness, identifier accuracy, abstention on ambiguous results, p95 latency, cost variables, and reviewer time. Add adversarial fixtures for prompt injection, malformed records, rate limits, and stale schemas. Publish failures with the same visibility as successful summaries.
For team adoption, add signed manifests, scoped credentials, policy-as-code checks, and an audit log that separates raw evidence from generated interpretation. A future host adapter should pass the same contract and security suite before it is marked compatible.
What to imagine next
A mature ecosystem could exchange evidence bundles between agents, attach uncertainty and data lineage to every claim, and let a reviewer approve a specific citation rather than an entire opaque answer. Skills could advertise cost and latency budgets, while databases publish schema-change notices that automatically pause affected workflows.
Those ideas remain a roadmap, not promises of the current repository. The practical next step is to prototype one narrow bundle, test it with real reviewers, and contribute improvements that keep the contract inspectable and reversible.
如何选择
| 比较维度 | 方案 A | 方案 B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
实施步骤
- 1
Choose one domain, adapter, model, and public fixture.
- 2
Store inputs, identifiers, raw hashes, citations, drafts, and review state.
- 3
Add adversarial tests, signed manifests, scoped credentials, and policy checks.
- 4
Replay across hosts and contribute only changes that preserve the contract.
可复制示例
{
"skill_revision": "<commit>",
"query_hash": "<sha256>",
"identifiers": [],
"sources": [],
"raw_response_hash": "<sha256>",
"draft": "<separate from evidence>",
"review": { "status": "pending", "reviewer": null }
}常见问题
Is a portable evidence bundle part of the current standard?
Treat it as a proposed project direction. The current repository provides skills and host guidance; a bundle schema still needs implementation and review.
What is a good first contribution?
Add a focused fixture, clearer provenance fields, safer credential boundaries, or host-compatibility tests for one existing skill.
资料来源
- Scientific Agent Skills README (captured 2026-08-31)来源核查 2026-08-31
- Scientific Agent Skills repository来源核查 2026-08-31