~/cxo.dev/blog/things-are-getting-interesting
← Notes
Weekly notes · · Claire Vo + Zach Davis

Things are getting interesting

Three emerging enterprise AI trends—meta-harnesses, collaborative agent work, and job-specific agents—are making the next phase of adoption more tangible.

A lone observer stands beneath a balloon-mounted observatory above a moonlit sea of violet clouds, with coral dawn on the horizon.

I mean, it's not like things have been uninteresting up until now. It's just that there are several emerging trends that make us feel very excited about the future of AI in the enterprise. From where we sit, for the conversations we have with companies every week, it really feels like the tooling is starting to catch up and things are about to get real interesting.

Multi-player agent work

We've long been fans of working out in the open. We tell everyone who will listen to build or buy an agent that runs in the cloud and can be at-mentioned in Slack. This is the number one unlock to cross-function, widespread adoption of AI agents. It's amazing for awareness. It unlocks collaborating with your teammates and agents at the same time. It fosters learning. We could go on and on. Lots of the most switched-on teams have already figured this out, some building custom solutions, but the tooling support just keeps getting better and better.

Slack announced first-class agentic coding support thsi week and is leaning hard on the "working in the open" angle. Block dedicated a whole closing section in their Berd blog post to their recently released collaboration platform Buzz. And we talked to an AI-native startup last week that is building their own custom harness with working in the open as one of its core tenets (everyone can see everyone else's sessions) and superpowers. For anyone who's been feeling isolated by the solo nature of most modern agentic workflows, it's time to check again. Collaboration is back.

Hiring agents for a job

The first time we taught a course together Claire put "Hire your agentic teammates" on one of the slides and Zach rolled his eyes. But Claire was right. She said it then and it's still true today:

To be successful your agents need to:

  • Live where your teams live
  • Have a defined job
  • Get access to the tools they need

With recent releases like Grok Bot, Block's Berd, and Slack Code, it seems like other people agree. We don't think the generalist agent is going anywhere, but the job-specific agent seems to be gaining steam.

The meta-harness

We wrote a little about this last week after Spotify announced Xirp and it just keeps coming up. Several companies we've talked to recently have built custom software and infrastucture that wraps harnesses and makes it painless to switch between them between (or even during) tasks. This is different from a custom harness that replaces Claude Code or Codex; this is a wrapper that encourages using whatever harness you're most comfortable with or is best for the task at hand.

This week we saw Block and Warp also talk about building a layer on top of coding harnesses. No doubt the great harness wars will rage on, but we're watching carefully because we think the meta-harness may start taking hold (but it might need a better name).

Here's what we're reading (and watching) this week.


Grok Bot + Grok 4.6 + Cursor Origin — is Claude Code dead?

Claire put a trio of recent Grok and Cursor releases through the ringer this week. We're both really liking Grok Bot and even replacing some of our existing agents with it. Grok 4.6 held it's own against the other frontier models. And Cursor Origin, a would-be GitHub killer, maybe needs a little more time baking in the oven.

Introducing Slack Code: Agentic coding for teams

Slack is giving collaborative agentic coding first-class treatment with this latest addition. As we mentioned up top, this solidifies a few trends we're big fans of: working with agents out in the open, and interacting with agents in the same places you interact with other teammates.

When a team works with AI out in the open, context compounds. Better context means better output, which makes AI more useful for the next person, and the next.

Warp Factories

Today, we introduced Warp Factories: flexible, open infrastructure for companies building cloud software factories. Moving from interactive agents to automated development in the cloud is a big shift, but it's pretty clearly the right one. Here's why

One thing we like about Warp's approach to the factory bandwagon is that they're pretty clear-eyed about what factories are and aren't today. They say they're automating about 30% of their tasks through their factories. That's a lot compared to the 0% that anyone was automating 6 months ago, but a far cry from the 90+% that the factory hype train would leave you to believe. As usual Zach Lloyd breaks complex agentic coding ideas in a no-nonsense way that's easy to understand.

Designing AI with character: what we learned building Berd

Block released Berd this week. It's a desktop application that wraps harnesses, bringing consistency and persistence even when switching between Codex, Claude Code, or their own open source harness goose. And like the recently released Grok Bot, they've leaned into the "personality" of agents, making it easy to customize not only the functionality but also the appearance of individual agents.

Introducing fx, a tiny, open, native coding agent from Vercel Labs

One final release of note: Vercel open sourced fx, a minimalist coding agent designed to feel like a unix tool. They say it was driven by three principles: fast, light, and open.

A single native binary, no runtime to install. It cold starts in 10µs and does no unnecessary work or I/O before accepting input. fx is the answer to "how fast can a coding agent be?"

After Codex and Claude brought one of our brand new Macbook Pros to a screeching halt last week by simultaneously gobbling up too much memory, "single-digit megabytes of memory at baseline" also sounds like music to our ears.


One thing to try this week

What else? Hire yourself an agent.

Pick one recurring responsibility you'd like to delegate. Write down a clear job description like you would for an unproven human assistant: what the job is, what context and tools are needed, what a good outcome looks like, and when to escalate to you. Give it to Grok Bot, or Codex, or Claude Code, or whatever you have available and ask for help getting it set up.


Stay interesting and interested.

— Claire + Zach

Get the next issue in your inbox at CXO.dev Notes.

Let's go.

If you're stuck, we want to help. Tell us what's working (and what's not) and we'll get you where you need to be.