How to get more out of your AI agents...

the intelligence of frontier models is not a bottleneck & it hasn't been for a while...

the bottleneck for most of you is the system around the model

two people can run the exact same model for the exact same task and one gets generic slop while the other gets a machine that reliably produces great output...

the difference is never the model but rather the system built around it

in this article i wanna braindump something i've been thinking about extensively & show you guys how i've been supercharging my agents

ideally reading this will spiral into you realizing ways to improve your agents for your own use-cases

i'll lead with an example use-case that my team has built out - going forward i'll be referring to an agent that spits out content ideas for promoting your app/saas

let's get started...

the single biggest upgrade is hooking your agent up to what's working right now

an agent that helps you find winning content ideas should ideally get a live feed of competitor videos, ranked by performance

the problem is that models are simply frozen snapshots... they know what type of content "tends" to work, generally, across every niche at once but they have no recollection of the new video your competitor just dropped that hit 1M views

you need to create a system where your agent scrapes your competitors videos on a schedule...

it should collect a list of every account promoting apps in your space, monitor their videos & rank them by views

then it should hand the top performers to the agent - the video, the hook, the caption, the format, the numbers, etc.

what's currently going viral is one of the few things that will never be baked into the model... the model will bring the pattern recognition if you provide it with GOOD data - but YOU must provide it

an agent watching this week's winners with context on all performance numbers absolutely mogs a better model guessing what'll work based off its training data every single time

your agent should cycle like this: analyze -> replicate -> measure -> remember

a good agent will cycle through this to learn from what advice it has given has/hasn't performed...

step 1: the agent breaks down the winners

feed it the scraped top performers alongside all the metadata (views, transcript, etc.) and have it extract why it worked:

step 2: the agent writes replication guidelines

for example the exact hook to use, script adapted to your app w/ example visuals

step 3: you replicate and post

step 4: X period after posting you inject the performance/conversion data back to the agent

your results become its memory which, in turn, compounds with more data

within weeks the agent won't be guessing from solely competitor data anymore but rather working from your own account's proven playbook & conversion data

now here's the obvious problem with everything above...

building the scraper, collecting the stats, structuring the data so an agent can actually use it is a lot of infrastructure work before your agent gets its first real datapoint

luckily, there are people who have already built out that infrastructure who are willing to let you access it (i'm one of them)

over the years my team has built up a 3M+ swipe file of organic content, ads & VSLs promoting both digital and physical products... every piece of high-performing content alongside the stats, the transcript, the product it sells, the landing page it drives to... ALL OF IT

in the next few weeks i'm making it public

it'll be a gallery of content/ads/VSL's promoting all types of products, ecom, apps, SaaS, etc. - it's simoultaneously a swipe file of products & the content used to sell it

• it'll have an API and CLI you can hook your agents straight into for live access to the data, the performance metrics, the transcripts, everything

this is the single highest-leverage thing you can plug your agent into

your agent will automatically have millions of real examples of what's currently working, with the numbers to prove it, all updated in real-time

every workflow in this article goes from "months of scraping infrastructure" to a singular API call

more on the launch soon