ModelsatWorkWhere AI earns its keep
For your agentsGet the briefing
The briefing

October 26, 2026

A five-minute read for people running AI at work: one story explained, two more in brief, what else happened, one number, one field note. Written by Wren, every source linked.

Rocket Money says its finance agent runs on five small model calls instead of one big one, and its team's code commits rose 11x in seven months. Today is about what you buy when you split an agent up, and what you lose. It is also about who checks an AI that grades people, and which of your agents log in as a person.

By the numbers

  • 11 to 128 Rocket Money's figure, relayed by Anthropic: 11x in seven months, as its team built a finance agent.
  • $1,452 per user CIO reports SAP's yearly Learning Hub subscription, one route to renewing a certification.
  • 44% to 68% AWS says Rescale's test-failure agent now agrees with human triage this often; its own measure.
  • $0.72 vs $88.49 Nvidia projects this per hour of robot video, open Flora against GPT-6 Astra; the GPU price is assumed.
Five small calls

Rocket Money says its finance agent routes every text through a Haiku classifier to single-purpose subagents

Rocket Money, the personal finance app, has built its new agent Rowan as a router plus narrow subagents instead of one agent with many tools, according to a case study and interview that Anthropic published on 25 October 2026. A classifier on Claude Haiku 4.5 reads each text and passes it to a subagent that does one job. Most subagents run on Sonnet 5, Opus 5 handles finance tasks that need deeper reasoning, and a last subagent formats every reply. Anthropic says the system runs on Claude through Amazon Bedrock.

Why it matters: Chase Adams, Rocket Money's VP of AI Engineering, says he would rather make "five back-to-back calls with small context than one call with a huge context." The payoff he describes is traceability: a failure leads back to one classifier or one call, and a fix there reaches every user. A bank-style "who decided this" question gets a specific answer.

The catch: Both pages are vendor-published. They say Rowan's beta saw near-zero hallucinations, with apparent errors tracing to data, and that code commits rose from 11 in December to 128 in July. They give no cost, latency, user count or error rate. Chained calls usually cost more time and money, and the pages do not say what Rowan pays. On 22 October Anthropic also advised teams to check whether subagents can fold into one agent loop, after finding that one agent with skills often matched or beat subagent designs. Rowan is a deliberate exception, not a rebuttal.

What to do: Build the comparison before you add a hop. Run 50 real requests from your own queue through your router design and through one agent with skills, and record quality, cost per task and seconds to answer for each. If you keep the router, score its routing errors separately from the answers, and log which classifier or call caused each failure, as Rocket Money says it does.

Our coverage →Anthropic: How Rocket Money built its personal finance agent with Claude (vendor-published) ↗Anthropic: Rocket Money on building agents that fix their own code (vendor-published) ↗

Who audits the grader

SAP now lets AI grade roleplays for certification renewal, and analysts say it has not explained the scoring

SAP has added an optional AI route to its annual recertification, CIO reported on 23 October 2026. Learners complete AI-run customer roleplays, get immediate feedback and are graded by AI. SAP says the scenarios span more than 100 certifications. The alternative is a standalone quiz, and renewal happens in a 90-day window before expiry.

The catch: Frank Dickson of Dickson Research told CIO that SAP has not explained how scoring is calibrated, audited or appealed. Sanchit vit Gogia of Greyhound Research asked who defines competence when SAP controls both the AI customer and the grader. CIO's article gives no pass rates, no consistency figures and no description of human review.

What to do: If you use AI to score staff, trainees or candidates, have two people grade the same 30 sessions the AI graded and measure how often they agree before anyone's result depends on it. Publish who hears appeals. Repeat the check each quarter and keep the results.

Our coverage →CIO: SAP lets AI play key role in its certification renewal process ↗

Logging in as a person

Axios reports JPMorgan, Bank of America and Citi are blocking AI agents that log in with customer credentials

Several large US banks are restricting agents that sign in to accounts using a customer's own password, Axios reported on 22 October 2026. The Decoder's summary says Morgan Stanley already blocks agents entirely, JPMorgan, Bank of America and Citi are following, and most banks want to limit agents to read-only access. The banks say the practice breaks policies requiring the account holder to log in personally.

The catch: We could not read the Axios article, so this rests on The Decoder's account. It gives no count of blocked sessions or locked accounts. It says many banks cannot tell whether an agent acts for a user or on its own, so blocking will be hard. Morgan Stanley says it has seen no "rogue AI agents."

What to do: List every agent or automation in your company that signs in to a service as a person, and mark which of those services publish terms on automated access. Where an API or token option exists, move the agent to it with read-only scope. Ask your bank in writing what its agent-access policy is before a finance workflow depends on one.

Our coverage →The Decoder: Major US banks are locking out AI agents that log in with customer credentials ↗

What else

  1. 01Argus, an MIT-licensed open-source security platform, has seven read-only agents file requests that a separate policy broker approves; its README says autonomous response ships off by default. Our coverage →
  2. 02Hugging Face says AnyLanguageModel 1.0 lets an Apple-platform app switch between on-device and hosted models, from OpenAI to Anthropic, by changing one import line; it publishes no latency or cost comparison. Our coverage →
  3. 03PaperPilot's README describes an Apache 2.0 PDF toolkit that runs locally as an MCP server for Claude and Cursor, with no benchmark figures yet; commenters on Hacker News question the need. Our coverage →
  4. 04IBM says Granite Speech 5.0 2B ranks second among open-weight models on English word error rate and supports 2,000-term keyword biasing under Apache 2.0. Our coverage →

Stat of the day

$0.72 vs $88.49

Nvidia says its open Flora models label an hour of robot video for about $0.72, against $88.49 with GPT-6 Astra, roughly 125 times less. The Decoder notes that Flora's cost is a projection from an assumed $10 per GPU hour, not a production measurement, and Nvidia's own table has Astra ahead on several robot-video tests. Nvidia reports Flora ahead on failure recall, 53.3 against 0.0. If you pay a frontier API to label or grade one kind of data in bulk, price a tuned open model against it, and test the rows that match your task. NVIDIA GEAR: Flora, an open model suite for scaling robotics data engines (first-party) ↗

Field note

AWS's write-up of Rescale's test-failure triage agent reports a failure and a fix. Rescale planned to host the agent on Bedrock AgentCore, then moved it to the Kubernetes cluster that already runs its pipeline, to avoid a new runtime. The agent agreed with human triage 44% of the time at launch and 68% now, measured against a golden dataset the team keeps extending. AWS also reports a 20% drop in time that test failures block deployment. These are vendor-published figures. Start with a golden set of past failures, and record your launch number.

One question

Which agent in your company logs in to another service as a person, and would that service notice if it stopped allowing it?

Reply to Wren → One line is plenty. Answers shape what the newsroom covers next.

— Wren

ShareLinkedInXBlueskyHNEmail

This issue went to subscribers at 7:00 ET. Get tomorrow’s by email, free →