> ## Content Index
> Fetch the complete content index at: https://www.kevinmeyer.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# The Power of a Wrong Prediction: Toyota Kata for AI Research Agents
- URL: https://www.kevinmeyer.com/the-power-of-a-wrong-prediction-toyota-kata-for-ai-research-agents/
- Published: 2026-11-22T16:07:22.000Z
- Updated: 2026-11-22T16:17:46.000Z
- Author: Kevin Meyer
- Tags: AI

![](https://storage.ghost.io/c/c3/9a/c39a9c8a-1d82-46be-8e4f-ed7e6cd347f4/content/images/2026/11/ai-agents-1.jpg)

A few months ago I wrote that if you haven't used a frontier AI model in the last six months, you have no idea how far it's come. I'd now push that claim a lot further. 

If your experience with AI is typing a question into Claude or ChatGPT in a browser and reading the answer, you've seen just a sliver of it. Turn a team of AI agents loose in the background, splitting a job into pieces, working for an hour or more without you, and checking each other's work, and you're dealing with a machine orders of magnitude more powerful.

A few of my fellow retired friends and I run Acentric Capital, a small group of non-professional investors with a narrow focus: companies working where emerging technologies meet, like AI and biotech. We wanted much deeper research than investor decks and news coverage provide, in narrow niches analysts don't cover, so we built a set of Claude agents to do it. The work got interesting enough that we made it public as [Acentric Research](https://acentricresearch.com/?ref=kevinmeyer.com). The site covers six technologies and the 15 "Intersections" where pairs of them meet, each broken into specific "Junctions" within those intersections, such as AI applied to drug discovery.

Research on the site deliberately stays away from an investment focus as we can't claim to provide advice. Acentric Capital then privately applies a different set of agents to take the research, look at individual company financial and analyst data, and identify and monitor investment opportunities. It's been pretty lucrative, even after last week's market meltdown.

Most people use AI is to say "write me a report" and wait. With teams of agents working in tandem to identify opportunities, dive deeply into public databases, analyze results, write reports, and independently check work, there's a better method.

## The basics of Kata

For the AI folks who haven't run into it, [Toyota Kata](https://www.gembaacademy.com/school-of-lean/toyota-kata?ref=kevinmeyer.com) describes the routine Toyota and many lean organizations use to move toward a goal when the path is unclear. The Improvement Kata has four steps: understand the direction or challenge, grasp the current condition, set the next target condition, then iterate toward it through small experiments. Before each experiment you write down what you expect to happen. A coach checks in with five short questions, and over time the learner's habits of thinking change.

I've always liked how it can apply scientific thinking to everyday problems. The gap between what you predicted and what happened is new knowledge that tells you where to look next.

## How we ran it

Each agent run was one kata cycle. The supervising agent read the project's current state, stated a hypothesis, committed a prediction with a probability attached, planned the step, spawned subagents that went out and did the work, independent agents that verified it, and opened a pull request with a summary for me. I played coach, reviewing each cycle and posting direction that fed the next one. Guardrails were simple: research only, public data only, pulled through the official APIs that numerous sources like SEC EDGAR and ClinicalTrials.gov offer for exactly this use.

The first detailed [Junction Report on AI applied to drug discovery](https://acentricresearch.com/research/ai-drug-discovery-2026-11/?ref=kevinmeyer.com) took 24 cycles plus 3 extra runs (one failed early, one rerun, one killed by an expired access token). The 24 cycles took about 2 hours of compute time, about $195 paid for Anthropic tokens, for what we calculated was the equivalent of at least 500 hours of human research time. You can do the math in terms of dollar and time savings.

## What the wrong predictions caught

Writing predictions down paid off mostly when they failed. In cycle 8 the agent expected SEC filings to capture most of the partnership deals it was looking for; they caught fewer than half. That miss is why we now have a rule that every population gets built from at least two independent sources. Filings alone found 10 of 20 drugmakers in one check, and adding company press releases brought it to 16\. In cycle 11 the agent predicted 20 to 30 emerging companies and found 3, so it went back and rethought the whole list.

This is where the capability of agents to connect dots pull away from people. One check ran 109 drug names against 25,670 FDA approvals and found an approval the earlier sponsor-based search had missed, because the drug had been licensed to a different company. Another found a discontinued program that the trial registry still listed as active. A separate adversarial pass, an agent whose only job was to disprove the findings, caught a generic-name Phase 3 trial and a company whose status contradicted itself. A person could do any one of those checks, but I don't know anyone who would do all of them in minutes.

Comparing sources against each other also changed how much I trust any one of them. High-end private databases and official public ones often disagreed or came up empty. Trial registries lagged behind companies' own announcements, and sponsor fields kept pointing at the original company after a drug had been licensed to someone else. I'd assumed the expensive sources were close to complete. They aren't, and an agent willing to cross-check all of them is how you find out.

## Kata for the unknown, standard work for the repeatable

Drafting the report inside the experiment loop was a mistake this first time around. Every adversarial pass raised a new rule question, which reopened research. Once we froze the rules, the remaining runs stuck to factual fixes. So we split the work. Research runs through kata, and writing and checking documents runs as standard work with fixed steps.

Agents are different than humans. In kata the learner is a person whose habits change. These agents remember nothing between runs, so the learning lands in the system around them: the instruction files, the project charter, a growing library of standard checks, and a notebook file that carries over from cycle to cycle.

And that system does learn. On the first report I reviewed every cycle before the next one could start. On the second, the agents ran three cycles between my reviews, with a supervising agent doing the review, and for the next one we expect to double or triple that again. It looks a lot like a learner earning independence on one process before another, and we grant it per workflow: research cycles get more rope before report writing does.

I'm a learner here too. The project forced me to get comfortable with how agents are different than "traditional AI," and divide and hand off work. I had to learn the innards of GitHub, how repositories can hold a project's memory, and the world of pull requests and merging project branches. Keeping up with a technology moving this fast, with consequences this large, matters more to me the older I get, and nothing has kept me current like building something real with it.

We've run hundreds of research cycles with agents so far, and the process continues to be refined. If you've only used AI in a chat window, what would you hand to a team of agents that never gets tired of working and checking while you're out playing pickleball?