AgentForce For All

AgentForce For All
Create a new Tableau Viz in Tableau Desktop

Inspiration

Tableau Next is more than just some new features, it represents a deeper integration and overlap between Tableau's core suite of products and the Salesforce platform. In this hack I wanted to take that integration one step further, to leverage the API first nature of AgentForce, Data Cloud and Tableau Semantics in order to bring Concierge to Tableau Desktop.

What it does

Thanks to the Authoring SDK pilot which was first debuted at Tableau Conference 2025 Devs on Stage (Labs Edition) I was able to do two things:

The extension is able to read the workbook metadata and find the Tableau Semantic model ID used by the Tableau Semantics connector. This is important as it can be passed to AgentForce to give it context about what data to query.
The authoring SDK allows the extension to write back to the workbook, in this case taking the viz metadata returned by AgentForce and translating it into a a Tableau viz.

These two capabilities allow AgentForce For All to answer questions of the data referenced by the workbook in exactly the same way as in the Tableau Next interface while also allowing the analysis to be saved to Tableau Desktop and used in a future dashboard.

Challenges I ran into

Coding against the response from an AI agent is difficult. From the same question Concierge will return one of a metric, a payload of data to draw a chart or a description of which fields to query directly in the semantic model and there is very little you can do to determine which response you'll get. When a metric or a payload of data is returned it can be more difficult to determine how to replicate the result in Tableau Desktop (i.e. know which fields to use in the viz) and so sometimes creating a new sheet just isn't possible. What makes this difficult for development is once AgentForce has a cache of data from previous questions it seems to not want to go back to the semantic model and therefore limits how often a new Tableau worksheet can be created.

Accomplishments that I'm proud of

Building my first every AI interface
Hacking the Agent's variables to get it to work with a semantic model rather than just a viz, metric or dashboard
Achieving something that a few people at Tableau have told me over the past 6 months just would never happen (that response always makes me want to just go off and build it myself)
Getting to know the Salesforce platform's APIs

What I've learned

Coding with agents is tricky but it can be a lot of fun and also never used the AgentForce or Data Cloud APIs before so learning how to write Tableau Semantic SQL was pretty cool. Mainly though I'm appreciating more and more how extensible both Tableau and Salesforce are opening the doors to a new level of possibility for anybody to create new features and solutions.

What's next for AgentForce For All

Depends whether Matt Miller can get the green light to make the authoring SDK available for general use.

I realise that I've not been able to provide source code or a SDO for the judges to be able to test it out themselves and so that probably disqualifies me. I'm happy to be available for a screen share or a call with any relevant PM to discuss the innovation further. I'd really like to see Tableau core users be able to benefit from what I'm increasingly appreciating is a powerful and flexible, ready-to-go AI data analyst.