Ratel

Fighting
context rot
relentlessly

[Ratel @ratelai]

Six months after shipping your AI agent, you don't really know how it works anymore.

The tool list grew and someone added instructions and nobody remembers why.

The agent is slower and more expensive.

This is what we call context rot.

There's a free test that shows you exactly what's rotted.

1. npx skills add ratel-ai/skills --all
2. /ratel-assessment

[Ratel @ratelai]

@Giac_nicoli 🦡🧡

[Giac @Giac_nicoli]

Yeah that was us 🦡🧡

The internet is now monetizing it…

So we open sourced it :)

Link in the first comment 👇

[Ratel @ratelai]

@0xCodez we want more things like this, gj

[Ratel @ratelai]

@lightsilver323 So i don't need a loan to pay my agents bill :)

[Ratel @ratelai]

@BatsouElef building the context layer for your agent so you don't need a loan to pay anthropic

[Ratel @ratelai]

@mikenevermiss @drewiustinai

[Ratel @ratelai]

@Star_Knight12 Not professional level yet but with tools like HeyGen you can reach a good result :)

[Ratel @ratelai]

rate your agent 1 to 10.

i expected an 8.
it scored a 5.8.

turns out giving it 100 tools wasn't smart.
i wouldn't want to read 100 pages every time someone asked me something either.

curious what you'd score.
npx skills add ratel-ai/skills --all

[Ratel @ratelai]

@rstagi_ @merybenavente Next merch coming!

[Ratel @ratelai]

Spreading romance while engineering context 🧡🦡

[María Benavente @merybenavente]

… and they say romance is dead.

[María Benavente @merybenavente]

… and they say romance is dead.

[Ratel @ratelai]

@merybenavente 😊

[Roberto Stagi @rstagi_]

merchmaxxing at the @aiDotEngineer 🦡

@ratelai @flaratel

[Ratel @ratelai]

We came to SF to interview the best minds at the AI Engineer World Fair.

Dario Amodei and daddy Jensen were busy.

We interviewed him instead...
Not that insightful tbh :(

Who should we talk to next?

[Ratel @ratelai]

Claude Opus 4.7 used 81% fewer tokens with Ratel.

Accuracy stayed within 1.7 points.

A local model went from 8% → 77% accuracy.

Fable we are coming for you!

#AIAgents #ContextEngineering #LLMBenchmarks

[Ratel @ratelai]

8% → 77% accuracy.

Same model and laptop.

With the full tool catalog visible, the model got overwhelmed.

With Ratel, it only saw relevant tools per task.

Better accuracy and 57% fewer tokens.

Less noise = Better agents.

[Ratel @ratelai]

Stop wasting tokens on bloated tool catalogs.

Ratel loads only the tools your agent needs, when it needs them.

Progressive disclosure for agents :)

[Ratel @ratelai]

pnpm add @ratel-ai/sdk

[Ratel @ratelai]

50 tools in your agent = 6000 tokens per run before it does anything.

We tracked it with a production client: Ratel's filter cut token usage 40%.

That's ~$1800/month, ~$22k/year back. Per agent.

Book Seychelles :)