On August 24, 2026, Caddi announced an "agent that builds agents" for back-office work in wealth management, law, and insurance. The product is not positioned as a prompt box. It first reads the systems a firm already uses, surfaces repetitive processes, and helps turn them into runnable, auditable agents.
Caddi says the product identifies repeated work and ranks processes by frequency and cost, allowing operations leaders to choose what should be automated first. That step matters because many back-office tasks look simple from the outside. Filing a document, for example, can hide version checks, decision branches, and exceptions that only the most experienced employee knows.
In a design canvas called Loop Studio, the person who performs the job can share a screen and demonstrate the process. Caddi says the agent does more than record the steps: when it reaches an ambiguity, it stops and asks, and the expert's answer becomes a rule in the process. The approach moves exception capture into design and training instead of discovering every edge case after deployment.
Caddi also says its agents use AI where judgment is required and deterministic code elsewhere. The company says each run can be replayed and that leadership can see every agent in one place. The release also claims SOC 2 Type II compliance and connections to more than 100 business applications. These are vendor descriptions in the launch announcement, not independently verified performance findings.
The release names customers in investment advisory, legal, and professional-services settings and describes workflows involving transitions, inbound mail, and conflict checks. Those customer examples and efficiency statements also come from Caddi's release and should not be treated as universal benchmarks.
The product direction is notable because it moves the automation bottleneck from whether a team can write code to whether it can see the real process, capture exceptions, and preserve an accountable chain. In back-office work, the hard part is often not the happy path but the dozens of exceptions. Without version control, permission boundaries, approval points, and recovery paths, multiplying agents can make governance harder rather than easier.
Caddi's announcement is a clear signal for the enterprise-agent market: the next layer will include systems that discover, train, and manage agents, not only agents that execute tasks. Whether this approach remains reliable across different data, permissions, and exception rules still requires deployment evidence and independent evaluation. For now, it is best read as a product direction for structuring process knowledge.
