Introduce AWS Strands Agents hook to common AI provider and Skills integration#67450
Introduce AWS Strands Agents hook to common AI provider and Skills integration#67450gopidesupavan wants to merge 3 commits into
Conversation
| def tool_fn(*args: Any, **kwargs: Any) -> Any: | ||
| return fn(*args, **kwargs) | ||
|
|
||
| tool_fn.__name__ = spec.name.replace("-", "_") |
There was a problem hiding this comment.
spec.name.replace("-", "_") only handles hyphens. Names with spaces ("my tool"), dots ("ns.tool"), leading digits ("123-tool" -> "123_tool"), or empty strings still pass through and produce invalid Python identifiers, which breaks Strands' inspect.signature()-based schema inference downstream.
Suggest a stricter normalizer plus a non-empty guard, e.g.:
import re
safe = re.sub(r"\W|^(?=\d)", "_", spec.name) or "tool"
tool_fn.__name__ = safeor reject invalid names up-front in ToolSpec validation.
|
|
||
| def _skill_spec_to_native(self, skill: str | SkillSpec) -> Any: | ||
| """Convert a skill source to a Strands-native skill object or path.""" | ||
| if isinstance(skill, SkillSpec) and not skill.path: |
There was a problem hiding this comment.
Precedence behavior here silently drops fields: when a SkillSpec has both path set and inline name/description/instructions set, the not skill.path branch is skipped and super()._skill_spec_to_native(skill) returns skill.path -- the inline fields are discarded with no warning.
Also, validation only happens inside the Strands branch, so a SkillSpec(name="x") (missing instructions/description, no path) only fails when Strands is the backend. Other hooks via the base path get a different error or none.
Recommend moving the mutually-exclusive validation into SkillSpec.__post_init__ so misuse fails fast regardless of backend, and explicitly rejecting path + inline combinations rather than silently preferring one.
| ) | ||
|
|
||
| def test_connection(self) -> tuple[bool, str]: | ||
| """Validate the connection by instantiating the model (no API call).""" |
There was a problem hiding this comment.
Minor: the "(no API call)" claim isn't guaranteed. get_model() calls GeminiModel(**kwargs), whose constructor behavior depends on the strands-agents/google-genai SDK version -- some versions probe credentials or metadata at construction.
Either drop the parenthetical or pin it: "no remote call as of strands-agents X.Y". Otherwise a future SDK bump could quietly turn test_connection into a network call without anyone noticing.
| - /docs/apache-airflow-providers-common-ai/operators/llamaindex_embedding.rst | ||
| - /docs/apache-airflow-providers-common-ai/operators/llamaindex_retrieval.rst | ||
| tags: [ai] | ||
| - integration-name: Strands Agents |
There was a problem hiding this comment.
Missing how-to-guide: for the Strands Agents integration. The Pydantic AI and LlamaIndex entries above each list how-to-guide: pointing at their .rst files (see lines 47-50 and 58-60), but the Strands block has only integration-name, external-doc-url, and tags.
Without how-to-guide, the new strands.rst / example_strands.py won't be linked from the integration index page in the docs. Add the .rst paths once the docs file is in place.
| prompt="Extract tables from report.pdf and summarize the findings.", | ||
| llm_conn_id="strands_default", | ||
| system_prompt="You are a document processing assistant.", | ||
| skills=["/opt/airflow/skills/pdf-processing"], |
There was a problem hiding this comment.
Hardcoded host paths (/opt/airflow/skills/pdf-processing here, /opt/airflow/skills/web-research further down) won't exist on a stock install -- anyone copy-pasting the example will hit a confusing failure at agent build time.
Suggest one of: (1) use Variable.get("strands_skill_path") so the path is configurable, (2) point to a path that ships with the provider (e.g. relative to example_dags/skills/), or (3) add an inline comment explicitly telling the user to place their skill bundle at that path before running the DAG.
| self.output_type = output_type | ||
| self.toolsets = toolsets | ||
| self.skills = skills | ||
| self.skills_params = skills_params or {} |
There was a problem hiding this comment.
skills_params was added to template_fields (line ~148) but is stored and used as a dict. Jinja renders a dict to a str (e.g. "{'strict': True}"), not back to a dict, so at execute time dict(self.skills_params) either raises or yields character pairs.
Options:
- drop
skills_paramsfromtemplate_fieldsif templating it isn't a real requirement, - register a JSON renderer via
template_fields_renderers = {"skills_params": "json"}and pre-parse, - or template individual values rather than the whole dict.
Same caveat already applies to the pre-existing agent_params -- worth a follow-up there.
2 participants
Add Strands Agents hook to common AI provider
Summary
Add AWS Strands Agents as a new agent backend for
AgentOperatorand@task.agentin the common AI provider, building on theBaseAIHookcontract.Introduce
StrandsHook(shared Strands SDK integration) andStrandsGeminiHookas the first backend (conn_type:strands-gemini, default connection ID:strands_default)Wire toolsets through
_tool_spec_to_native, convertingToolSpecinstances to Strands-native toolsAdd skills support end-to-end:
SkillSpecdataclass onBaseAIHook,skills/skills_paramsonAgentRunRequestandAgentOperator, and StrandsAgentSkillsplugin integration for filesystem paths and inline skill definitionsRegister the new connection type in
provider.yaml/get_provider_info.pyand add optional dependency:pip install 'apache-airflow-providers-common-ai[strands]'(
strands-agents[gemini]>=1.0.0)Add example DAGs (
example_strands.py) covering basic operator usage, skills, inlineSkillSpec+ SQL toolset, direct hook usage, and@task.agentDocument connection setup, hook usage, and operator skills in new/updated RST pages
Depends on
BaseAIHook PR #67438
Follow-ups
Durable execution for Strands (
durable=True)StrandsHookcurrently setssupports_durable=False. A follow-up PR should mirror the pydantic-ai durable path so Strands agents can resume from cached steps on task retry.Out of scope for this PR: usage limits for Strands hooks.
Skills for Pydantic AI (
pydantic-ai-skills)PydanticAIHookcurrently leavessupports_skills=False, soAgentOperator.skills/skills_paramsonly work with Strands backends in this PR. A follow-up should wire the same operator-level skills API to pydantic-ai via thepydantic-ai-skillslibrary (Agent Skills / agentskills.io spec with progressive disclosure).Files changed
hooks/strands_ai.py,hooks/base_ai.pyoperators/agent.pyexample_dags/example_strands.pyprovider.yaml,get_provider_info.py,pyproject.tomldocs/connections/strands.rst,docs/hooks/strands_ai.rst,docs/operators/agent.rst, …tests/unit/common/ai/hooks/test_strands_ai.py,test_base_ai.py,operators/test_agent.pyWas generative AI tooling used to co-author this PR?
{pr_number}.significant.rst, in airflow-core/newsfragments. You can add this file in a follow-up commit after the PR is created so you know the PR number.