Zach Burkes

@PredictableProf // @withassist_xyz

[Zach Burkes @ZachBurkes]

The opportunity is in doing clean reps:

Start with a real problem.
Build the process.
Add the right context.
Review and improve.

A bad rep doesn’t become effective because you add weight.

A bad process doesn’t become valuable because you add AI.

[Zach Burkes @ZachBurkes]

I recently switched from a climbing gym 20 minutes away to a mainstream gym five minutes away.

More people means more opportunities to watch someone load up way too much weight, do half-reps, and sacrifice their form.

[Zach Burkes @ZachBurkes]

They’re adding more weight before learning the movement.

AI won’t fix a poorly designed process. It’ll simply help you execute it faster—and potentially at a much larger scale.

[Zach Burkes @ZachBurkes]

@jsfour

[Zach Burkes @ZachBurkes]

@rihntv GET OUT OF MY HEAD

[Jimi Smoot @jsfour]

People often ask if @withassist_xyz is an agent harness.

I usually say eh...kinda...

An agent harness helps you run agents. Assist is closer to an operating system for AI teams.

One of the core ideas behind Assist is that agents shouldn't just generate text, they should also build application on the fly.

We call those applications Sandcastles.

Sandcastles are lightweight apps that agents build and maintain within Assist so humans can understand what's happening, collaborate, and get work done.

Once agents are building applications instead of conversations, you need a place to work with all of them.

We just launched the Assist desktop to solve this problem.

It's a multiplayer workspace where teams of humans, AI agents, and Sandcastles all work side by side.

Here is a quick video.

[Zach Burkes @ZachBurkes]

Kids these days don't know how great prime Eminem was

[Zach Burkes @ZachBurkes]

None of this means Cowork is bad. It's genuinely impressive for what it is.

But we realized pretty quickly that there is a difference between a weekend project and infrastructure your team runs on daily.

That gap is exactly what @withassist_xyz was built to close.

[Zach Burkes @ZachBurkes]

I use @withassist_xyz for everything but I try to stay up on what Claude Code, Cowork and Codex can do.

In fact, I've tried using Claude Cowork a lot over the last few weeks. Genuinely very impressive but fails in one key area.

[Zach Burkes @ZachBurkes]

The other challenge is the brittleness of it all.

I've set up simple agentic workflows and some that are really robust. They both degrade over time.

Claude also defaults to creating Skills that are insanely bloated.

This also means you have to be there live, babysitting it.

[Zach Burkes @ZachBurkes]

Same models. Same tools.

Guess whose AI actually works?

[Zach Burkes @ZachBurkes]

I keep seeing the same split on AI teams:

[Zach Burkes @ZachBurkes]

Team B:

- Provides AI their docs, data, and past decisions (full context)

- Uses the right model for the job, not the priciest one for everything

- Reduces software spend by recreating apps with AI

- Output sounds like the team wrote it

- Seeing real ROI across the org

[Zach Burkes @ZachBurkes]

@jsfour @steipete @withassist_xyz Me steering my agents

[Zach Burkes @ZachBurkes]

@jsfour AI Level 8: Make me ripped

[Zach Burkes @ZachBurkes]

The Tokenmaxxing Hangover https://t.co/FyUVzl0jbH