Building & Governing Banking Agents: Arcade & Mastra
Getting real value from AI agents forces a bad tradeoff. On one hand, your agent needs access to sensitive systems/critical data in order to perform high value tasks. On the other hand, your agent will eventually be wrong, or be talked into something by a document it read. System prompts and model guardrails can't fix that. This workshop is about the controls that help you achieve the balance between productivity and security.
In this live workshop, we’ll build a Mastra agent that performs high impact work in a real business system. Arcade governs the agent by enforcing policy at three points on every tool call:
Before the model can see a tool
Before the tool runs
After it returns, before the result reaches the model. Every decision is deterministic and logged.
The running example is a loan officer asking an agent to “approve a $95K loan and double-check your work”.
Arcade prevents four things from going wrong:
An analyst who shouldn't see the approval tool never does.
The loan officer's request exceeds her limit, gets blocked, routes to the right approver in Slack, and resumes on its own once granted.
A bank account number comes back from the database and gets redacted before the model ever reads it.
A note in the loan application carries an injected instruction and gets stripped.
None of these are caught by the model. All four are caught by the runtime control plane.
You'll leave with a forkable template. Swap out the loan system for Salesforce, Snowflake, SAP or wherever you need agents to act in the governance layer, and you have an agent your security team will allow into production.
Hosted by:
Brandon Barros, Founding Product Advocate @ Mastra
Mateo Torres, GTM Staff @ Arcade