Rocket Money's Rowan agent: a Haiku router, narrow subagents

Rocket Money, the personal finance app, has built its new agent Rowan as a router plus narrow subagents rather than one agent with many tools, according to a case study and an interview that Anthropic published on its site. Users reach Rowan by text message and ask it to check spending, negotiate a bill or cancel a subscription. Both pages are vendor-published and were read on 25 October 2026.

How Rowan is put together

The case study says a classifier running on Claude Haiku 4.5 reads each request and routes it to a subagent built to do exactly one thing. Most subagents run on Sonnet 5. Opus 5 is used for financial tasks that need deeper reasoning. A final subagent formats every reply so Rowan's voice stays consistent. It says the system runs in production on Claude through Amazon Bedrock.

Chase Adams, Rocket Money's VP of AI Engineering, says in the interview: "I would rather have five back-to-back calls with small context than one call with a huge context, because of recency and primacy effects." Aaron Dignan, VP of Product and AI, says the bar in finance is 100% accuracy and that a failure can be traced to the exact classifier or call that caused it.

Routine work becomes code; the agent repairs it

Rocket Money calls its approach "third way engineering". Adams describes it this way: a cancellation is codified as a procedure that runs repeatedly until it breaks, usually because a website changed. At that point an agent works out what broke, fixes the procedure and turns it back into code. He says Agent Skills, which can include scripts, matched what the team was already doing.

Controls they describe

• Structured output, so evals check deterministic results rather than asking a model to judge itself.
• Classifier fall-throughs that catch off-topic requests early.
• A Claude check on whether an agentic action, such as a cancellation, actually completed; if it did not, the agent classifies why and hands off to a human to decide what happens next.

What the numbers say

The case study reports near-zero hallucinations during Rowan's beta, with apparent errors tracing to data rather than the model, and an 11x rise in monthly code commits, from 11 in December to 128 in July. Those are Rocket Money's figures as relayed by Anthropic. They do not say how many users or conversations the beta covered.

The catch

• No cost or latency figures are given, and those are the usual price of chaining several calls.
• On 22 October Anthropic published guidance recommending teams check whether subagents can fold into one agent loop, after finding in several enterprise deployments that one agent with skills often matched or beat subagent designs on quality, cost and latency. It also says subagents still fit verification, parallel work and tool-selection problems. Rowan is a customer's deliberate exception, not a test of that advice.
• Rowan's free-trial tracking, which reads Gmail, is described as about to roll out, so it is not yet evidence of anything.

What to do

• If you run a router plus subagents, log which classifier or call caused each failure, as Rocket Money says it does, so one fix reaches every user.
• Test your own design against a single agent with skills before keeping the extra hops; measure quality, cost and latency per task.
• Put a cheap model on routing only if you can evaluate its routing errors separately from the answers.