Gareth Andrew retweetledi
Gareth Andrew
1.5K posts
Gareth Andrew
@gingerhendrixai
Software/AI Engineer. Hacker and experimenter. Vibe Coder.
Belfast Katılım Ağustos 2024
812 Takip Edilen366 Takipçiler
Gareth Andrew retweetledi
Cloudflare Dynamic Workers are FAST! I'm running 450 unique scripts/second with a p99 runtime of ~22ms 🤯
Kenton Varda@KentonVarda
Dynamic Workers are now in Open Beta, all paid Workers users have access. Secure sandboxes that start ~100x faster than a container and use 1/10 the memory, so you can start one up on-demand to handle one AI chat message and then throw it away. Agents should interact with the world by writing code, not tool calls. This makes that possible at "consumer scale", where millions of end users each have their own agent writing code. blog.cloudflare.com/dynamic-worker…
English
Gareth Andrew retweetledi
new toys:
- codemode: chew through multiple tool calls in one/few shots
- shell: local fs, that you can back with do/d1/any sql db(!) and r2. git support incoming.
- worker-bundler: package npm deps and serve full apps (!)
do you see what's happening.
Cloudflare@Cloudflare
We’re introducing Dynamic Workers, which allow you to execute AI-generated code in secure, lightweight isolates. This approach is 100 times faster than traditional containers. cfl.re/4c2NvPl
English
Gareth Andrew retweetledi
Kenton and Sunil cooked so hard on this one
not sure we're ready for the 2nd order downstream effects
Kenton Varda@KentonVarda
Dynamic Workers are now in Open Beta, all paid Workers users have access. Secure sandboxes that start ~100x faster than a container and use 1/10 the memory, so you can start one up on-demand to handle one AI chat message and then throw it away. Agents should interact with the world by writing code, not tool calls. This makes that possible at "consumer scale", where millions of end users each have their own agent writing code. blog.cloudflare.com/dynamic-worker…
English
👀 the piece i was waiting on
@cloudflare/worker-bundler" target="_blank" rel="nofollow noopener">npmjs.com/package/@cloud…
English
imo, the best thing Cloudflare have built since durable objects
it's a beautiful primitive but needs a lot of user-land love, I predict a lot of cool things are going to get built with this in the next few months.
Cloudflare@Cloudflare
We’re introducing Dynamic Workers, which allow you to execute AI-generated code in secure, lightweight isolates. This approach is 100 times faster than traditional containers. cfl.re/4c2NvPl
English
which coding agent has the best harness?
i don't mean pricing, so don't just say "cAuDE"
English
I keep meaning to do a proper write up, but simplified it's just using the filesystem for all context e.g.
agent 1: write a file [research]
agent 2: read file, and update (or write a new one) [plan]
...
agent X: read file(s), write code [implement]
agent Y: read code and file write new file [review]
I still have some chats but only for brainstorming, and they start with some other agents writing a "context bundle" so the chat agent doesn't waste tokens going off exploring by itself. (I actually plan to turn off tools entirely for chat but not quite there yet)
English
@gingerhendrixai I'm intrigued. what exactly does this mean in practice?
English
Conversations are dead (to me), and it's working out well.
Gareth Andrew@gingerhendrixai
I think it's time to admit conversational threads are an anti-pattern. Every message should be able to exploit the full context, especially the smart fraction. There is never a situation where appending a message to a trajectory is the best we can do.
English
YouTube
English
watching a (very good) @dexhorthy talk, but can't shake the image of a noble warrior trying to teach the villagers how to fight before the invading armies arrive
English
sadly no prizes for being first, but Letta were way ahead of the game here
letta.com/blog/sleep-tim…
eric zakariasson@ericzakariasson
you can try this out in cursor today! cursor.com/marketplace/cu… this will look at your conversation history every N prompt, spawn a subagent to extract memories, and then store them where the agent can access them easily. if you're curious, there's a full article in thread!
English
I've named this iteration of my personal system "streams", after a walk in the hills after heavy rains had me pondering similar metaphors.
recommended read:
Thorsten Ball@thorstenball
Is software turning into a liquid? Is this what most software is going to be? Nameless, shapeless? Created (poured?) just in time, evaporating just after? My current fascination in this week's Joy & Curiosity intro. registerspill.thorstenball.com/p/joy-and-curi…
English
@LambrosPetrou @badlogicgames @opencode Pi gives you tonnes of scope for customisation. Though I am trending the other way - building light abstractions over a pluggable harness so I can drop in CC or codex or pi etc
English
I want to build my own "very lightweight" harness to do certain "deep research" like automations.
What's the best harnesses to use? I am exploring Pi from @badlogicgames and I know @opencode supports what I want too but it's heavyweight.
Others, with integration for Go or RPC?
English
@kentcdodds You could do worse than point your agent at mastra's OM
English
I think I'm liking the direction of a personal AI assistant built using MCP and Cloudflare Dynamic Worker Loaders (code mode). I love that I can talk to it through ChatGPT, Claude, or Cursor (existing subscriptions). I just need to build out a good memory system (biggest risk).
English
Gareth Andrew retweetledi
@gingerhendrixai Yeah it's good for parallelism and context management. But the puppet show approach isn't really useful, it's just something people do because they're used to thinking about people.
English
Agree. There are some technical reasons (around context management) that justify multi-agent setups, but in most cases they are actually not that beneficial. Better future models will not need that.
Tom Goodwin@tomfgoodwin
I’m surely being stupid. But if AI is rather unconstrained by expertise or capacity or to some extent speed Why do we need to divide tasks or departments to 9 agents ( the marketing agent, the optimization agent etc ) to each do one thing. And then another agent to manage the swarm. Cant one agent just be doing it all you know. It seems very skeuomorphic. Will we have HR agents to make sure the agent agents are being looked after ? A office canteen manager agent to feed the agents ? Seems daft
English
Jarred now has anthropic and OpenAI stock!
Jarred Sumner@jarredsumner
Astral was my first (& so far only) angel investment. A small check in their Series A.
English