docs: add pydantic-deep to third-party toolsets #3747
Conversation
| - `TodoToolset`: Task planning and progress tracking for agent self-organization | ||
| - `FilesystemToolset`: File operations (ls, read, write, edit, glob, grep, execute) | ||
| - `SubAgentToolset`: Spawn and delegate tasks to specialized subagents | ||
| - `SkillsToolset`: Load and use [agent skills](https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills) from markdown files |
There was a problem hiding this comment.
Assuming (some of) these can be used stand-alone, without the rest of the Deep Agent framework, would it be possible to extract these into separate packages, that we can consider for upstreaming once they've matured a bit?
If they are usable standalone, even if they're still in the same package, I also wouldn't mind subheadings so that it's easier for a user to find "Todo Toolset" or "File system Toolset" in the sidebar ToC if they just want that bit.
There was a problem hiding this comment.
Maybe it makes sense to have the explanation of your framework in the multi-agent-... page, and then here have a separate section for each of the stand-alone-usable toolsets, that can say "Part of ..." linking to the Deep Agent doc.
Those sections should be more generic, less marketing copy, just saying things like "File system access is useful because blablabla. Check out these implementations by the community:", as they may have multiple links, e.g. also https://pypi.org/project/pydantic-ai-filesystem-sandbox/.
There was a problem hiding this comment.
Thanks for the suggestion. I've analyzed the toolsets for extractability:
Easy to extract:
TodoToolsetSkillsToolset
Would require more work:
FilesystemToolset- tightly coupled to theBackendProtocolabstraction layerSubAgentToolset- dynamically imports other toolsets and is coupled toDeepAgentDeps
I'll look into extracting TodoToolset and SkillsToolset as separate packages first. Need to think through the API design and whether to keep them compatible with pydantic-deep or make them fully independent. Will follow up once I have a plan.
In the meantime, I can add subheadings for each toolset in the docs for better discoverability in the sidebar.
There was a problem hiding this comment.
I'd like to have headings under "Third-Party Toolsets" (after MCP Servers, before Langchain) for each type of common toolset that has one or more community implementations and then in that section list each option, i.e. "Task Management"/Todos", "Filesystem" (https://pypi.org/project/pydantic-ai-filesystem-sandbox/ + yours once it's available), "Memory", "Code Execution" (could list https://github.com/pydantic/mcp-run-python), etc.
The ones that Pydantic-Deep covers can then have an item saying something to the effect of "The third-party Pydantic-Deep deep agent framework includes a FilesystemToolset among others"
That way it'll be easier to discover for users, not being nested under the not-necessarily-related-seeming Deep Agents heading, and leaves room for more future community alternatives.
The copy on the Multi-Agent page can then be the main place where that concept is introduced + Pydantic-Deep is mentioned more clearly, as I don't think a "Deep Agents/Pydantic-Deep" heading necessarily makes sense on the Toolsets page. I also wouldn't be opposed to a new page like https://starlette.dev/third-party-packages/ listing Third Party packages explicitly; that's also where https://github.com/jagreehal/pydantic-ai-guardrails could go for example.
| - **Context management** — automatic conversation summarization to handle long sessions that would otherwise exceed token limits | ||
| - **Human-in-the-loop** — approval workflows for dangerous operations like code execution or file deletion | ||
|
|
||
| These patterns emerged from teams building production agents (like [Claude Code](https://docs.anthropic.com/en/docs/agents-and-tools/claude-code/overview) and [Manus](https://manus.im/)) and discovering what actually works at scale. |
|
|
||
| These patterns emerged from teams building production agents (like [Claude Code](https://docs.anthropic.com/en/docs/agents-and-tools/claude-code/overview) and [Manus](https://manus.im/)) and discovering what actually works at scale. | ||
|
|
||
| [Pydantic-Deep](https://github.com/vstorm-co/pydantic-deepagents) implements these deep agent patterns for Pydantic AI. See the [toolsets documentation](toolsets.md#pydantic-deep) for usage examples and the [full demo application](https://github.com/vstorm-co/pydantic-deepagents/tree/main/examples/full_app) for a complete example with a chat interface, file uploads, skills, and streaming responses. |
There was a problem hiding this comment.
I want to make it clear that each of the items in the list above can already be implemented using some Pydantic AI feature / 3rd party package, (we should have links to each doc, e.g. the toolset sections you're gonna add), and that if one doesn't want to piece this together oneself, Pydantic-Deep is a community package that offers an opinionated take on a deep agent by leveraging these same features.
There was a problem hiding this comment.
I'm happy to take a crack at this at the end to make it sound a little bit more from-Pydantic than from-VStorm.
| - `TodoToolset`: Task planning and progress tracking for agent self-organization | ||
| - `FilesystemToolset`: File operations (ls, read, write, edit, glob, grep, execute) | ||
| - `SubAgentToolset`: Spawn and delegate tasks to specialized subagents | ||
| - `SkillsToolset`: Load and use [agent skills](https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills) from markdown files |
There was a problem hiding this comment.
I'd like to have headings under "Third-Party Toolsets" (after MCP Servers, before Langchain) for each type of common toolset that has one or more community implementations and then in that section list each option, i.e. "Task Management"/Todos", "Filesystem" (https://pypi.org/project/pydantic-ai-filesystem-sandbox/ + yours once it's available), "Memory", "Code Execution" (could list https://github.com/pydantic/mcp-run-python), etc.
The ones that Pydantic-Deep covers can then have an item saying something to the effect of "The third-party Pydantic-Deep deep agent framework includes a FilesystemToolset among others"
That way it'll be easier to discover for users, not being nested under the not-necessarily-related-seeming Deep Agents heading, and leaves room for more future community alternatives.
The copy on the Multi-Agent page can then be the main place where that concept is introduced + Pydantic-Deep is mentioned more clearly, as I don't think a "Deep Agents/Pydantic-Deep" heading necessarily makes sense on the Toolsets page. I also wouldn't be opposed to a new page like https://starlette.dev/third-party-packages/ listing Third Party packages explicitly; that's also where https://github.com/jagreehal/pydantic-ai-guardrails could go for example.
|
|
||
| The framework also includes multiple backends (in-memory, filesystem, Docker sandbox), automatic conversation summarization, and human-in-the-loop approval workflows. | ||
|
|
||
| See the [documentation](https://vstorm-co.github.io/pydantic-deepagents/) for installation and usage, or explore the [full demo application](https://github.com/vstorm-co/pydantic-deepagents/tree/main/examples/full_app) for a complete example with a chat interface, file uploads, skills, and streaming responses. |
There was a problem hiding this comment.
The path has a risk of getting outdated, so let's not include it.
2 participants
Summary
Adds pydantic-deep to the third-party toolsets documentation. Pydantic-deep is a framework built on pydantic-ai that provides ready-to-use toolsets for building autonomous agents with planning, filesystem operations, and multi-agent capabilities.
Changes
toolsets.md: Added Pydantic-Deep section under Third-Party Toolsets with:
multi-agent-applications.md: Added Community Frameworks section with reference to pydantic-deep
Affiliation: Vstorm