Ananta
Reads what you asked for, writes it down word for word, and sets what “finished” means before any work starts.
Who does what
18 agents, each with one job. None of them can sign off on its own work.
Command
Take the order and run the job
Brahma
Breaks the job into steps, hands each one out, and judges every result that comes back before choosing the next move.
Build
Design it, feed it, make it
Mayasura
Plans the structure of bigger systems, such as data layouts and how the pieces connect, before they're built.
Visva
Writes and runs the code in a sealed workspace, to a fixed budget of attempts.
Vyasa
A silent daemon that collects and cleans raw data so every job starts from real inputs.
Check
Prove it before anyone trusts it
Agni
Reviews code for errors and risky behavior before it's trusted to run on real data.
Ketu
Tries to break things on purpose, so weak spots are found before customers find them.
Sanjaya
The numbers core. Measures results instead of giving opinions about them.
Kalki
Watches results over time and flags anything that stops performing the way it did.
Guard
Stop what shouldn't happen
Nandi
A hard-coded watchdog with no AI in it. It can stop any job, and no model can talk it out of that.
Vasanti
Holds authority to override the supervisor when something crosses a line.
Kubera
Tracks every dollar spent on compute and services, so no job quietly runs up a bill.
Aksh
Runs simulations on live conditions and raises the alarm early.
Remember and publish
Learn from it, then ship it
Saraswati
Keeps the permanent record of past failures so the same mistake isn't made twice.
Narada
Posts finished, verified work to each platform. Nothing unapproved goes out.
Markets desk
Specialists for market work
Ram
The final link between the swarm's decisions and the systems that act on them.
Ekalavya
A self-taught model that looks for patterns the others can't see.
Priyanka
Reads the human side of the market: mood, news, and momentum.
The rules they work under
Each rule exists because something went wrong without it.
"Done" comes from the check
An agent can say it's finished. The job only closes when the files, columns and row counts are read from disk and match the order.
Your order is kept word for word
The original request is attached to every step, so a short follow-up like "keep going" can't quietly drop your filters.
Real inputs or none
If a source file is missing, the job stops and says so. Agents never build a stand-in and pass it off under the real name.
A run that finishes instantly is wrong
A job that "processes" thousands of rows in half a second gets treated as a failure and investigated.
Loops get cut off
If an agent rewrites the same file four times and the check is still failing, it has to stop and hand over a summary.
Budgets are fixed before work starts
Every job has a ceiling on steps, time and spend, and rented GPUs are shut down as soon as a batch is done.
Good tools get kept
When a job passes, the working tool is saved and approved, so the next client's run is faster and cheaper.