Tim Smart

twitch.tv/timmeeeeeeeeh Working on clanka (effect native ai agents) and answering effect questions.

@kitlangton @thdxr A hipster and a homeless man walk into a bar...

I got some AI ceo insights before they hit LinkedIn

@skoshx Still an experiment, will run some proper evals soon comparing to normal tool calls.

@tim_smart is the "tools use through writing code" a long-term architectural decision or just a temporary experiment? from my intial testing it seems quite token inefficient, am curious to understand the direction!! super cool implementation wise though

Soon 👀 An Effect toolkit for building custom coding agents.

@BleedingDev It currently ships with support for openai and Copilot subscriptions

@tim_smart Will it be on top of existing subs and CLI tools and SDKs? Because honestly that's the appealing part - with subscriptions.

@3ddelano Yep. You could have the harness in the browser connected to an executor running elsewhere.

@tim_smart Will it be possible to have this run the agent the browser and have it access the local file system via proxies

@lancewharper @imax153 It helped with some things :) But lalph now has clanka built into it.

@tim_smart @imax153 Lalph did this?

@StephenKin53884 AgentExecutor allows you to seperate where the actual work happens / tools calls from where the harness runs.

@tim_smart I’ve been building a durable, distributed agent runtime on top of Effect Cluster - e.g. a tool call is a cluster entity. Is this the kind of thing the AgentExecutor is for?

@tim_smart Built in support for subagents?

Live hacking on an Effect coding agent: twitch.tv/minbadblue

On a side quest 👀

What have I become...

Last week we launched Effect v4 Beta 🚀 If you missed the announcement, here’s a recap of the initial overview -- what changed under the hood, what to expect during the beta phase, and how to approach migration: youtube.com/watch?v=j7U4Qu…

@zeeg Span events are pretty good for errors. You can see the full stacktrace, any logs that were correlated as well as the full trace for context. But if you want to chat with the effect team we can setup a slack or discord channel.

Who do I know working on Effect that can help us fix the story for instrumentation so it correctly works with Sentry's SDKs? OpenTelemetry is not the solution. Its fine for traces and logs, not for errors.

@ethanniser Instead of Stream.rechunk(1), you probably should also try start the stream with single item chunks (instead of rewriting them, which is quite a bit of overhead)

did some benchmarking - and found effect v4 streams to be mostly on par with node streams, and in some cases faster depending on the workload node streams call 6 callbacks per-item 🤯 effects streams are more efficient and support chunking internally more details in reply

@waynemaurer @EffectTS_ @effect/atom-react" target="_blank" rel="nofollow noopener">npmjs.com/package/@effec…

@tim_smart @EffectTS_ I see it in the repo, but don't see it on npm? npmjs.com/search?q=%40ef…

Effect v4 is in beta. 🚀 Rewritten runtime. Smaller bundles. Unified package system. The most ambitious release we've made so far. Here's what's new. 🧵

All of this work has resulted in a leaner, simpler library that feels nicer to use. There are many other improvements in Effect v4 I could talk about, but for now I encourage you to give it a try and tell us what you think!

Also every other module has been given the same treatment. Instead of creating interpreters for everything (such as Stream or Schedule), modules now directly build upon Effect. Which improves both bundle size and performance. In some cases streams are 100x faster!

A year in the making! Excited to see people start playing with it. But how did we make Effect both smaller and faster? A thread 🧵