Langgraph/chains User Permissions

[levi-katarok]

Hi All! I'm trying to work on implementing a network of agents that could ideally scale to hundreds of agents. These agents need to be tied to specific user permissions. There are two primary challenges I'm facing:

1. Managing Hundreds of Agents:
   My initial idea is to decompose the network into multiple subgraphs, each with a primary assistant planner. This planner would then delegate tasks to cohorts of agents based on the use case. I'm curious if anyone has experience with this approach or can suggest a better strategy for managing a large number of agents.

2. Accessing User Permissions:
   Users should have access to specific agents based on their permissions. For instance, one user may have access to the flight booking agent but not the marketing agent, while another user has access to both.

I'm thinking that either the permission check occur at the invocation of each agent, or it could be the primary assistant agent filter and manage available agents based on user permissions. Ideally, I plan to manage user permissions in a database. Any insights on best practices for integrating user permissions with agent access would be helpful.

Has anyone approached a similar challenge? I think we're getting to a point where to be useful for companies deploying flexible user systems will be key.

[hwchase17]

off top of head, i would:

• pass in some user id as part of the config when invoking the agent
• in the planner, look up in a database which agents that user has access to
• only let the planner route to those agents

[levi-katarok]

Thanks for the prompt reply!

I'm going along a similar train of though. A couple ideas I'd like to try

• Dynamic binding of pydantic agent classes to the base agent as their tools. I'm thinking this way I can keep my entire graph intact and it just never knows that that's a tool available. And thus theoretically shouldn't be able to be branched to in the graph if that tool invocation is never triggered
• Maybe every agent is a conditional agent based on the user permissions.

Will update here.

[hinthornw]

Yeah definitely don't want to expose any distractions to the agent that would let it end up in invalid states. Anything you can exclude programmatically will lead to better outcomes

[vsdeepthi]

@levi-katarok were you able to figure this out? I have the same requirement of controlling access to Agents

[levi-katarok]

Hi yes, what I did was store sub agent to user permissions in a db. When I initialize my main agent runnable with a prompt and tools l only bind the the tools (actually sub agents) to the main assistant in which the user has permissions.

Hope that helps!

[vsdeepthi]

ok thanks Levi

…

On Fri, Aug 2, 2024 at 7:19 PM Levi ***@***.***> wrote: Hi yes, what I did was store sub agent to user permissions in a db. When I initialize my main agent runnable with a prompt and tools l only bind the the tools (actually sub agents) to the main assistant in which the user has permissions. Hope that helps! — Reply to this email directly, view it on GitHub <#1034 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABNGYOVS22QKEATQVJXVT23ZPOE6LAVCNFSM6AAAAABK63NU7KVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTAMRSGQYDSNY> . You are receiving this because you commented.Message ID: ***@***.*** com>

-- *Deeps,* *Reality is a story and we can choose our story. So why not pick "the better story"*

[vsdeepthi]

Is there a boiler plate code that you can post Levi?

…

On Fri, Aug 2, 2024 at 7:19 PM Levi ***@***.***> wrote: Hi yes, what I did was store sub agent to user permissions in a db. When I initialize my main agent runnable with a prompt and tools l only bind the the tools (actually sub agents) to the main assistant in which the user has permissions. Hope that helps! — Reply to this email directly, view it on GitHub <#1034 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABNGYOVS22QKEATQVJXVT23ZPOE6LAVCNFSM6AAAAABK63NU7KVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTAMRSGQYDSNY> . You are receiving this because you commented.Message ID: ***@***.*** com>

-- *Deeps,* *Reality is a story and we can choose our story. So why not pick "the better story"*