Seeking Gradient

Building for agents. lets your agents easily publish single page HTML apps (free!), and make them smarter over time (also free!)

[Seeking Gradient @seekinggradient]

I’ve been thinking about why AgentsFS feels powerful.

For models, test-time compute means the model spends more tokens thinking during the response.

For agents, the equivalent isn’t just “think longer when asked.” Instead, it’s doing useful cognitive work between requests: summarizing what happened, compressing raw material, resolving contradictions, writing durable briefs, cleaning up stale notes, and leaving the next agent a better starting point.

That changes the shape of “memory.”

A lot of agent memory today is either hidden memory in the agent harness, raw chat history, or bigger context windows. Useful, but incomplete.

Agents don’t just need more context. They need better future context.

Frontier models get better by spending more tokens reasoning at inference time.

Agents get better by spending more time improving the environment they’ll infer from next time.

That means turning messy inputs like docs, transcripts, or failed attempts into concise, accurate, source-cited, up-to-date notes the next agent can actually use.

https://t.co/Nabcx4aCiE is just a very simple to understand, completely local, lightweight implementation of a filesystem where agent knowledge can compound.

[Seeking Gradient @seekinggradient]

I’m going to do a follow up post later this week talking about some of the stuff I’ve been building and why sprites is so important

[Seeking Gradient @seekinggradient]

https://t.co/W5ACfeiRtn might be the single biggest unlock to building your own distributed agents.

Sprites is a game changer.

I spent this weekend listening to every YouTube video they’ve produced. Love it when a company puts this much time into creating great content.

Sprites is a great product pivot, and I’m so glad they did it

@flydotio

[Seeking Gradient @seekinggradient]

CodexBar is honestly great. It's also a misnomer.

The Mac Menu Bar also shows your Claude Code and Gemini usage in addition to Codex and has a bunch of goodies, including telling you how you are projected to be doing in terms of your rate limits.

[Seeking Gradient @seekinggradient]

2 tools I wanted to call out today from my weekend experimentation:

1. open-slide - https://t.co/WnYurYGd5V -- if you're building presentations, by far the best way to make them quickly is to have an agent go and build it using tools it's most familiar with, which in this case is HTML and JavaScript.

OpenSlide seems to be the best implementation of agent-native lightweight presentations. I had it create a slide deck to compare hotels I was considering for an upcoming trip and the slides came out great - still needed polish in some areas, but still remarkable.

2. https://t.co/qMgw83BOfU - comprehensive platform for generating AI videos. I saw a bunch of posts on X about building cinematic home pages, so I asked @grok what the best platform to create AI-generated videos is, and it said https://t.co/qMgw83BOfU is by far the leading platform.

It did not disappoint -- so many models hosted, and it was a breeze to setup.

[Seeking Gradient @seekinggradient]

Whoa, some weird system warning in Claude while using Fable across a couple compactions

"Model performance degradation detected"

[Seeking Gradient @seekinggradient]

Quick note to optimize your Fable usage: agents in Claude Code can spawn sub-agents that are less intelligent. A pattern that's working well for me is to have Fable act as an orchestrator of sub-agents that are using weaker models to orchestrate a large, complex task and maintain optimal architecture quality.

[Seeking Gradient @seekinggradient]

Really enjoyed @ishaansehgal's "The Log Is The Agent" talk. This seems super powerful -- the durable log is the real thing you own, while models/runtimes are replaceable interpreters. Feels important for portable, inspectable agents.

https://t.co/yirgLzv71v

[Seeking Gradient @seekinggradient]

Excited for @garrytan’s talk at the AI Engineer World's Fair tomorrow!!

[Seeking Gradient @seekinggradient]

Tried to tell Fable to slow down or pause, but it keeps working 😅. Didn't want to interrupt it

[Seeking Gradient @seekinggradient]

Used my 5 hour quota on Fable running a couple sessions through Fable on Ultracode

[Seeking Gradient @seekinggradient]

I love @flydotio 's blog.

Awesome https://t.co/U0ShVXW7ib

[Seeking Gradient @seekinggradient]

@oasisdevices Just ordered mine

[Seeking Gradient @seekinggradient]

Seems really cool

[OASIS @oasisdevices]

Today we introduce OASIS 1. 

The smart ring built for private dictation. Whisper to write. Touch to edit. 

A first step beyond the keyboard toward a world where your intent follows you across every device.

Order at https://t.co/gZieZw6vYJ first batch is limited.

[OASIS @oasisdevices]

Today we introduce OASIS 1. 

The smart ring built for private dictation. Whisper to write. Touch to edit. 

A first step beyond the keyboard toward a world where your intent follows you across every device.

Order at https://t.co/gZieZw6vYJ first batch is limited.

[Anthropic @AnthropicAI]

We’ve received notice that the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5.

We'll begin restoring access tomorrow, and will share an update soon.

We’re grateful to our users for their patience, and to everyone who worked with us on redeploying the models.

[Seeking Gradient @seekinggradient]

I might be late to the party here, but you might be too. One of the most trending conversations in AI/Agents I'm hearing right now are about VMs and Sandboxes.

Have you read this yet? You should: https://t.co/xEfdZdxdos

[Seeking Gradient @seekinggradient]

😂😂😂

[Trung Phan @TrungTPhan]

BREAKING: Lebron has made his latest free agency decision. He will join Anthropic.

When asked how many frontier models he plans to work on, the NBA legend said:
“Not one, not two, not three, not four, not five, not six, not seven...maybe 9 or 10.”

[Seeking Gradient @seekinggradient]

Honestly, dynamic workflows are unreasonably effective and all you need.

[Seeking Gradient @seekinggradient]

I have a local Qwen model running in ollama on Pi right now, and it keeps making typos again and again and is getting frustrated.

Some reason, in a SQL query, it keeps typing `NOMALIZED` instead of `NORMALIZED`, and it's getting mad at itself in the thoughts