AI Research Agents Use Toyota Kata to Cut 500 Hours

Team of AI research agents running Toyota Kata experiment cycles

A small investing group says AI research agents running on a Toyota Kata routine did in about two hours what would take a person at least 500 hours. The group, Acentric Capital, built the setup on Claude agents and published the output as Acentric Research.

The detail that matters most is how the agents were made to fail usefully. Before each run, they wrote down a prediction. The misses are where the learning came from.

How the research agents were run

Acentric Capital is a handful of retirees investing as amateurs in companies where AI, biotech and other new technologies overlap. Their public site maps six technologies, the 15 pairings between them, and narrower topics inside each pairing, such as AI in drug discovery.

Toyota Kata is a habit lean manufacturers use when the route to a goal isn’t clear. You name the challenge, describe where things stand, pick the next target, then run small experiments toward it. Each experiment starts with a written prediction.

Every run of the AI research agents counted as one kata cycle. A lead agent checked the project, proposed a hypothesis and attached a probability to its prediction. Subagents then did the digging, separate agents checked their work, and a pull request summarized the outcome. A person played coach, reviewed each cycle and set the direction for the next. Agents were limited to research on public data, fetched through official interfaces such as SEC EDGAR and ClinicalTrials.gov.

AI research agents image 1
Image: Kevin Meyer

Cost and speed

The first full report, on AI in drug discovery, needed 24 cycles and three more runs: one crashed early, one was repeated, and one died when an access token expired. The 24 cycles used about two hours of compute and roughly $195 of Anthropic tokens. By the group’s estimate, a human would need at least 500 hours.

What the wrong predictions caught

Writing predictions down paid off mostly when they failed.

  • In cycle 8 the agent assumed SEC filings would reveal most partnership deals. They revealed under half. Now every list has to draw on two independent sources at minimum. In one check, filings alone turned up 10 of 20 drugmakers, and press releases lifted the count to 16.
  • In cycle 11 it forecast 20 to 30 young companies and found 3, which sent it back to redesign the list.
  • Matching 109 drug names against 25,670 FDA approvals exposed one approval that a search by sponsor had missed, since the drug had changed hands through a license.
  • A different check spotted a canceled program the trial registry still showed as active.

A further pass used an agent assigned only to knock down the findings. It flagged a Phase 3 trial listed under a generic drug name and a company whose records contradicted each other.

The comparisons shifted how much the group trusts any one source. Premium private databases and official public ones frequently conflicted or returned nothing. Registries trailed company announcements, and sponsor fields still named the original company after a drug was licensed away. The paid sources, the group had assumed, were nearly complete. They were not.

Where the method broke down

Writing the report inside the experiment loop backfired. Every adversarial check raised a fresh question about the rules, and that reopened the research. After the rules were frozen, later runs only fixed facts. The team now splits the work: research runs through kata, while writing and checking documents follows standard work with fixed steps.

The agents keep no memory between runs, so what they learn is stored around them in instruction files, a project charter, a growing set of standard checks and a notebook that passes from one cycle to the next.

What comes next for AI research agents

Hundreds of cycles later, the process is still being tuned. Independence is granted one workflow at a time. For the first report a person reviewed every cycle before the next began. For the second, agents ran three cycles between reviews, with a lead agent checking. The team expects to double or triple that on the next report.

Research cycles get more rope than report writing does. Acentric Capital also runs a separate, private set of agents that apply the research to company financial and analyst data, though the public site avoids investment advice.

Source: Kevin Meyer

Written by

Saksham Bhargava

Saksham Bhargava runs Alteroid and has been writing about consumer technology for over five years, covering iPhone, Android, Windows, AI tools and gaming. With an MCA and six years of hands-on web development experience, he tests the features he writes about on his own devices and focuses on clear, practical guides that solve real problems. Away from the screen, he enjoys travelling to places most tourists skip.