Sakana AI gives agents a shared business playbook

Welcome to The Signal from Superpower Daily with Maya and Theo

What if I told you the biggest bottleneck in enterprise AI is not a lack of computing power

it is actually corporate jargon

Today we are looking at Sakana AI

They just introduced a new platform called Aquarium

It is trying to solve the massive headache of giving AI agents shared business knowledge

Well it is a

fascinating approach

I mean everyone wants autonomous agents right now

Nobody wants the massive set of nightmare that comes with them

Right

So let us start with the reality on the ground here

Enterprises are really struggling right now

They want to onboard these AI agents efficiently

They want these bots to handle finance and operations

And customer support

Exactly

But they are hitting a massive wall

They end up wasting all this time because they basically have to rebuild the business knowledge from scratch

for every single new agent Yeah And

that

is where Sakana AI is stepping in So they just introduced Aquarium They describe it as a managed foundation It basically

packages all that recurring setup work It turns

it into reusable infrastructure Which sounds

great on paper Right

I

mean think about what that setup work actually entails You

have business data connections You

have internal definitions You

have these really complex decision rules Sure

So

right now a company builds an agent for marketing team They wire

up all those connections Then the

finance team wants an agent And they

just have to start all over again Exactly They

have to wire up the exact same connections Aquarium creates

this infrastructure layer instead And this

layer can then be shared by multiple AI agents simultaneously

So this brings

us to a very specific technical concept Sakana relies heavily

on something called an ontology Which sounds like a word you would only hear in a university philosophy department We need to

break that down It does sound

incredibly academic But it is

essential for understanding how this actually works Think of an

ontology as a massive dynamic map It maps exactly

how

a specific company talks It maps the

relationships between

their different data silos It also maps their specific

decision criteria Let us use an analogy

here Imagine hiring a whole team

of human assistants at once If you did not have

a shared system well you would have to

write

a brand new company handbook for every single new hire Yeah I mean you would

sit them down one by one Right You would explain what

Q3 revenue means to your specific accounting team You would show them exactly

which server holds the client contracts It would just be exhausting

Exactly The ontology is that

ultimate

shared company handbook You write it once Aquarium

creates this shared map

Then all the different AI

assistants share it They all read from the exact same playbook So if the marketing agent

needs to know what a tier one customer is it checks the ontology Yes

And if the support

agent needs the exact same definition it checks the exact same

place But here is where it

gets really interesting technically Aquarium does not force you

to move your data That is the brilliant part

The data stays completely put

Enterprise companies absolutely hate moving

data It is risky It is expensive It breaks

compliance rules Oh

for sure So Aquarium

just leaves the

source data in its existing storage It works directly with your

SQL tables It plugs into your unstructured

documents It simply overlays the ontology

on top of what you already have It connects the dots without

moving the underlying files This fundamentally changes how workflows

operate You define a metric once

Every agent understands it perfectly

So workflows can now become

much more sophisticated They can combine different capabilities

seamlessly Right An agent can use

AI

judgment for a vague request And then it can execute

deterministic functions to pull the actual data And it integrates mandatory human

approvals That is a crucial detail here Oh that human element is

huge

Companies are terrified of letting AI

run completely wild For good reason I mean enterprise

environments are highly

regulated Aquarium accounts for this It maintains

a very strict audit

trail It records exactly what data the

agent used to make a decision It logs exactly who approved the

final action Yes It records the ultimate outcome

You

get full transparent visibility You

can see the execution speed

You can monitor the accuracy You

can even track the underlying

model usage So what does this actually mean

for you if you run an enterprise tech stack You might want to buy this

immediately But there is a catch You

cannot just swipe a credit card and access Aquarium today No you cannot The platform is

currently only available

through customer specific consultation Sakana has not announced any

public pricing They have not given a

general availability date either Sakana is taking a very

deliberate hands on approach They want

their customer

facing engineers working directly with organizations Building an ontology

is complicated They want to

ensure proper construction They want to

validate the results before releasing this to the broader public The pedigree behind

this product is definitely worth noting Sakana engineers have

serious enterprise backgrounds We are talking

about former Palantir engineers We have people

from Snowflake We have builders

from Microsoft and AWS They deeply understand

enterprise data pipelines They also have

specific financial sector experience Sakana cited past

projects

with major Japanese institutions They mentioned Mitsubishi UFJ

Bank They mentioned SMBC Group

They cited Daiwa Securities

Now we need to

be very precise about those banking connections Sakana cited those projects

as their foundational experience They are not new

aquarium deployments That is a very

important distinction Sakana engineers worked on

those specific banking projects in the past Those banks had AI

lending experts They had automated proposal

generators And the engineers saw

the pain of repeated setup work firsthand Exactly That specific pain

point

inspired the creation of Aquarium but those banks are

not currently running this new software It is simply the

history behind the product So what should you

watch next with Sakana Well there are features

still

in the planning phase They want the platform

to eventually update these ontologies dynamically because right now building

the map is a heavy lift They want the system

to learn from operational feedback That would be a

massive technical leap The system would essentially

rewrite its own handbook If an agent learns

a new relationship between two datasets the ontology updates automatically

Yeah every other agent

gets

smarter

instantly Sakana is also looking at

using AI to construct the agents themselves The infrastructure would basically build

the workers Okay that wraps up our

look

at Sakana's infrastructure Moving from a shared foundation

for multiple agents that naturally begs a bigger question If we are building these

massive systems do we actually need that

many individual agents running on them in the first place Right Which brings us to

Anthropic

Yes They just released new

October

guidance on agent consolidation And their conclusion is actually

quite surprising They state that fewer AI

agents can deliver better results and they also cost significantly

less This completely challenges the current

enterprise trend I mean the industry has

been totally obsessed with multi agent architectures Everyone has been

building these massive swarms of specialized bots Right now enterprise setups often

mimic

human org

charts It is almost comical Every department

builds their own agent

A marketing team builds a marketing

agent Customer support builds a support

agent Sales builds a sales agent

And then a platform engineering

team builds

a

master

coordinator agent

just to route tasks between them all It creates a corporate bureaucracy for bots It

really does You get the exact

same problems you

get with human bureaucracy You get overlapping rules You get conflicting

instructions Most importantly you

lose context as work

is passed from the coordinator agent to the marketing agent Nuances get dropped It is a digital

game of telephone So Anthropic

looked at this mess and

they decided to test a different

approach They actually recommend folding these specialist sub

agents into one main agent loop Yeah they found that a single

agent loop can execute tasks roughly twice as fast One customer measured massive speed improvements

just by ditching the multi agent swarm Which makes sense structurally A

single agent does not

have to spend time summarizing its work to pass it to the next agent It just does the work

Exactly Wait I have to

push

back

on this though If you take all the

detailed instructions from the marketing agent and the support agent and

the sales agent and you stuff them into

one single master agent aren't you gonna cause a

massive problem You would completely overwhelm its context

window The AI would just freeze up

from information overload You would think so If you

loaded everything into the

prompt at once it absolutely would fail The model would completely lose track

of what it was supposed to be doing But Anthropic is bypassing that limitation

They are using a concept called

progressive disclosure Progressive disclosure How exactly does that

work under

the hood So the main agent is not

given every single instruction up front It starts with a broad high

level overview of its job It sees very brief descriptions

of its available skills and tools It only loads the full

detailed

instructions when it actually needs them for a very specific step in

a workflow Oh I see It is like

a master chef

with a massive library of cookbooks They do not memorize every

single recipe before opening the restaurant for the night They just memorize the index

Right When an order for

a

grilled steak comes in they query the index they pull only the grilling

instructions into their active memory They

cook the steak then they clear their mind for the next order That is a perfect way

to visualize it Technically the agent is using system

prompts to dynamically retrieve instructions It only pulls them when

triggered by a specific task Retaining earlier results in one

continuous working history also helps immensely Because the agent can connect

findings much easier than separate bots can Exactly Progressive disclosure prevents the

context

window from overloading while keeping all the capabilities available But there are important limitations

here Faster is not automatically better

Anthropic ran a very specific

test with a support agent They folded a shipping specialist

sub agent directly into the main agent and it actually broke

the workflow The consolidated agent failed

one out of two times Which is a catastrophic

failure rate for an enterprise system It really is The

main agent just

got lazy It skipped checking a

required delivery date It just answered the

customer without consulting the complex shipping rules And that highlights the

massive risk of merging everything A faster answer can

easily skip a required complex step The model reaches for

the easiest path to a completion There is also a

hard technical limit on the number of tools you can provide Tool selection degrades rapidly

if you expose too many options So what is the

actual threshold Anthropic says you run

into trouble if you expose more than 30 50 tool definitions in the system prompt The model's attention mechanism just

gets scattered It starts picking the wrong tools So you cannot just give

one agent 100 tools and expect perfection It gets profoundly confused You absolutely

should retain sub agents

in specific edge cases Keep them if you hit

those context limits Keep them if parallel work

is needed If you need to search

a database and generate an image at the exact same time You need separate agents Keep them if tool instructions

heavily

conflict with each other The decision to

consolidate must be made

very carefully task by task So what should you

watch next Anthropic

recommends a very deliberate testing strategy

You should start consolidation tests with just

10 carefully chosen examples You must meticulously inspect which individual tasks

flip from pass to fail Do not just look at the overall

success score You have to look at the granular steps Anthropic even provides a migration kit on

GitHub

right now It helps engineering teams test this transition safely

It is a measured approach to simplification The

industry is realizing that more agents do

not automatically equal more intelligence Next up we are looking at how developers

are actually being heavily incentivized to build these exact types of models and workflows Meta and Cerebral Valley are launching a massive

10 day virtual AI hackathon It runs from October 26 to November

4

2026 It is backed by a 1 million prize pool And it is capped at 20

000 builders This

is a phenomenal push for grassroots AI

adoption Meta wants developers firmly entrenched inside its ecosystem

They are throwing serious resources at this to

make it happen Builders get immediate access to Meta's absolute newest

tools They get the Muse Spark series They get

Muse Wonder They get Muse

Code They get

access to SAM

3 1 for advanced segmentation And they also get

to use the Llama OpenWeights It is an incredibly well

equipped sandbox They are giving developers

the keys to the entire multimodal kingdom There is real money

involved to support the compute costs too Every selected participant gets

150 in MetaModel API credits There are also limited

partner credits available Base10 and Modal are

offering 100 each Fireworks AI is offering 25

The compute support is serious

Base10 even lists a 2

million token per minute rate limit for participants They are ensuring developers do not

hit artificial bottlenecks especially while prototyping complex agent loops I find the structure of this

competition fascinating There are two entirely distinct judging

tracks The first is the Indie or

Student track This track is all about exploration

It judges pure creativity It judges

technical ambition It basically

asks one main question

Is this a cool working idea

The second track is a completely

different beast It is the Startup or Founder

track This track demands rigorous execution It

looks for actual production readiness

It asks a much harder question

Can this product survive real world

users and scale gracefully They are essentially judging a prototype

versus a full product Both tracks use the exact same

themes though One theme is build for your

world The other theme is push the

frontier But the expectations for the final

submissions are completely different It is a very smart way

to allocate a million dollars You encourage the wild unconstrained ideas on one side We

also encourage serious investable business ventures on the other It

captures

the entire spectrum of developer

talent There are some major caveats to this

hackathon Entry is not automatic just because you

sign up It is a rolling selection process The

organizers look closely at your project

proposal They evaluate your technical experience You can

link your GitHub profile Right

You can link previously shipped products

Right

They are curating the list Once

20 000 qualified builders are

accepted the doors close permanently There

are

also very strict geographic

exclusions This is not open to the entire globe No it is not The official

rules bar applicants from

several specific countries This includes China It includes Russia

It includes Belarus

It includes Ukraine

It also includes

Syria They are

strictly enforcing these borders

And Meta and Cerebral Valley employees

are also entirely ineligible which is standard practice for these types of corporate events So what should you watch next

regarding this event The logistics are actually surprisingly unsettled

The official finalized rules are not

even published yet They will be released sometime before

the kickoff Organizers explicitly reserve the right to

change dates and times at their discretion Yeah The current event listing says

it

starts at noon Eastern but the exact kickoff live stream

time is still unconfirmed Applicants really need to watch

the event page very closely over the next few days It sounds like they are

still finalizing the infrastructure for 20 000 concurrent builders Meanwhile what happens when

the

tools

these developers build try to interact with strict enterprise security like your actual bank account

Right Axios reports that major

U

S banks are aggressively updating their terms of service Morgan Stanley Bank of

America and

Wells Fargo are completely blocking AI agents They are stopping these

bots from taking any autonomous actions inside customer accounts The banks are drawing

a very hard non negotiable line in the sand here They are strictly

separating the ability to read data from the ability to actually transact I have to

ask a blunt question here If I voluntarily

give an AI agent my login credentials shouldn't I be

allowed to use it It is my

money If I want

a bot to log in and pay my utility bills every month why is the

bank stopping me It comes down

to a massive liability and security nightmare for the banking institution Bank terms of

service explicitly require a human being to initiate a login If you hand

a bot your password you are violating

those terms Furthermore it creates

a

massive blind spot for their security systems Explain that blind

spot How does the

bank see it Well their web

application firewalls are designed to detect unusual behavior but a successful

login with correct credentials looks identical on the bank's end They cannot tell

if it is you typing at the keyboard They cannot tell

if it is a rogue AI agent executing a script Right So if

an

agent drains your account because of a hallucination or a prompt injection attack who is responsible

Under current regulations

the bank often

bears the burden of fraud They simply cannot

accept that uncontrolled risk That makes total

sense They are terrified

of automated high speed mistakes But they

are not blocking AI completely JPMorgan and

Citi do allow limited access They allow

view only data retrieval Yes

They use officially approved intermediaries like Plaid to handle this safely So

the AI can read your account balance It

can categorize your past expenses But

it absolutely cannot move your money Jamie

Dimon has spoken very clearly about this The

JPMorgan CEO publicly stated that unrestricted entry by outside AI agents is simply too risky He

said it is too risky for the clients He

said it is too risky for the bank's underlying infrastructure The

distinction is critical here View

only access is a window You can look inside Autonomous

action is handing over the key to the vault The

banks will not hand over the key Now

we do need to mention a major caveat regarding this Axios report The

specific incidents that sparked this reporting are actually unverified That

is a crucial detail Reuters

attempted to independently confirm the specific cases that Axios described They

could not verify them And

a Morgan Stanley spokesperson went on the record They

explicitly stated they have not seen any rogue AI agents accessing their platforms This

means we should not assume every major bank is currently under siege by automated bots The

policy changes are highly likely to be preventative measures They

are updating their legal terms before the technology becomes widespread They

are getting ahead of the problem So

what should you watch next in the financial AI space Keep

a close eye on the development of secure APIs Will

banks eventually approve official highly

secure API routes Routes

that give agents actual Heavily

guarded transaction authority Routes

that go beyond just view only access

Exactly

If they do they will likely require cryptographic signatures and distinct bot identities

rather than just shared passwords

It is absolutely the next major frontier of open banking

The demand for automated finance is just too high to ignore forever

We are now moving into our Quick Reads

I am ready

Anthropic and Theorem used Claude to find critical software bugs

These were bugs completely missed by years of conventional testing

Claude wrote 180 million lines of mathematical proofs These are

specifically called rock proofs

Claude based this massive effort on OSS fuzz tests across 123 different open source

libraries It successfully found a rare image decoder

bug This specific bug required a 64 megabyte

input It also required 2 gigabytes of memory just to trigger the

failure Standard fuzzing usually stops at 1 megabyte

inputs It just throws random data at a program to see if it

crashes Claude did

something entirely different

It translated the code into formal mathematical logic

It found flaws buried incredibly deep in the architecture

It even found non crashing logic flaws that traditional fuzzers mathematically cannot

see The caveat here is highly technical but

important The absolute guarantees of these mathematical proofs only apply to specific modeled

conditions In this case they modeled a RISC V processor running a Linux

operating system This does not mean the software is totally bug free in

every possible real world

environment The underlying math is

perfect but the real world of hardware and operating systems is incredibly

messy Uber is officially rolling out chat GPT work to all 30

000 employees globally This full rollout follows a highly successful

1

000 person pilot program Interestingly 80 of those early users were non

engineers

They were employees

in finance operations and product management These employees used

it to localize training materials rapidly

They built complex interactive daily dashboards Uber

also deployed 30 specialized AI engineers They organized them into two week AI pods specifically to hunt

for new automation opportunities across the company It is

a massive aggressive deployment of enterprise AI They

are deliberately giving access to

absolutely everyone not just the technical teams

They are essentially crowdsourcing the

discovery

of Of their own corporate inefficiencies

Exactly The caveat here is purely financial Uber has not disclosed the

actual cost of this massive global rollout Furthermore they have not

reported any quantified hard productivity gains yet They are deploying it specifically to see

what workflows emerge naturally from the workforce They

are closely watching employee adoption metrics first They will worry about measuring

the strict financial return on investment later AI stocks dragged Wall Street lower

this week The

NASDAQ dropped 1 46 overall Meta fell 6 7 They recently

issued non voting shares specifically to fund further AI

expansion Tesla fell 1 9 They recently cut their capital expenditure which raised

serious questions about the pace of their AI and robotics development Lam

Research tumbled 10 1 They reported a very slight profit forecast miss CoreWeave also fell

amid growing concerns about borrowing massive sums to

buy more NVIDIA chips The broader market is incredibly anxious right now High AI

valuations leave absolutely zero room for any kind

of disappointment or delay The real caveat here

is about shifting timelines Investors are increasingly treating financing needs as immediate stress

tests They are no longer treating them as long term visionary bets

The market's patience for heavy unproven AI spending is wearing incredibly thin The grace period is

over The bill for all this AI growth is

finally

coming due We are

now moving to three takeaways from today First infrastructure is steadily maturing We

see this clearly with Sakana building shared ontologies to standardize business rules We also see it with

Anthropic refining progressive disclosure to make single agents highly capable

without

overloading

The messy foundation of enterprise AI

is getting much cleaner Second big boundaries are rapidly being drawn Major banks are

strictly locking down agent autonomy to protect customer accounts from liability Meta is placing rigid

geographic limits on who can even participate in its

massive

hackathon The rules of engagement

for developers are tightening globally Third the market now demands proof Wall Street

is highly anxious over massive AI spending versus actual tangible returns Uber is deploying chat GPT to all employees

specifically to hunt for actual valuable workflows rather than assuming they exist