The model-agnostic approach is smart. Most teams I know are stuck between "we want AI" and "we can't send data to OpenAI." Being able to plug in local models while keeping the same workflow layer is the right call.
Curious how you handle tool permissions in practice. Is it just a whitelist, or can you scope down to specific rows/fields?
At the moment any tool that you write inline in your instructions - the agent has access to use.
The way to scope down access, would be to use an Automation (our RPA, zapier-like section). You can give an automation parameters and control exactly what it does, then use it as an agent tool.
In future, we'll be looking into user-delegate permissions. I.e. you give your users a role (RBAC) in Budibase, which gives them access to certain resources. Any agent calls will be done under the user's permissions. This is not always desirable - sometimes you want your agent to have more permissions that the user - so this will need to be configurable at some level.
The model-agnostic approach is smart. Most teams I know are stuck between "we want AI" and "we can't send data to OpenAI." Being able to plug in local models while keeping the same workflow layer is the right call.
Curious how you handle tool permissions in practice. Is it just a whitelist, or can you scope down to specific rows/fields?
At the moment any tool that you write inline in your instructions - the agent has access to use.
The way to scope down access, would be to use an Automation (our RPA, zapier-like section). You can give an automation parameters and control exactly what it does, then use it as an agent tool.
In future, we'll be looking into user-delegate permissions. I.e. you give your users a role (RBAC) in Budibase, which gives them access to certain resources. Any agent calls will be done under the user's permissions. This is not always desirable - sometimes you want your agent to have more permissions that the user - so this will need to be configurable at some level.