kevo

Co-founder at ContextStream where we’re building the best persistent context layers for AI coding agents.

[kevo @kevoscott]

Don’t get locked into an agent harness.

Create a project in ContextStream and use it as portable memory. Stash your Agents.md, soul.md, and user.md in docs, then pull them into whatever new harness you switch to.

After you install the new harness, connect ContextStream and instruct it to update with that info.

[kevo @kevoscott]

It’s wild that I learn to use my own product (contextstream) better by observing the cleverness of frontier models using the tools. What a time.

[kevo @kevoscott]

@dhh This is great. My kids think the screen savers are the coolest thing and always ask about Omarchy.

[kevo @kevoscott]

@simonw No. Almost exclusively chat -> voice mode

[kevo @kevoscott]

@ashxhart I’d make one of them ContextCode

[kevo @kevoscott]

@khemaridh We’ve had Team skills in ContextStream since April. Those skills show up anywhere the ContextStream MCP is enabled. Yes, quite useful.

[kevo @kevoscott]

@dhh @tobi Congrats!

[kevo @kevoscott]

I’ve started seeing advice now to have LLMs report the decisions made during an agent session.

I highly recommend this and it has been a core part of ContextStream from the very start.

We call it the knowledge stream. It’s a running feed of your decisions insights preferences, and lessons learned across a project. Its visible from the web dashboard for human review. You can edit it’s contents or the memory events themselves to adjust what stays in context.

However, unlike the list mentioned above, the agents get delivery of these memory events just in time through ContextStream. So there’s no required manual management of decision logs.

[kevo @kevoscott]

Best practice with contextstream, (although I think it could be applied more broadly) is to give a short explanation why with your instructions to the llm.

[kevo @kevoscott]

@RobbieHodge @BillDA Grok is the best search/ answer engine in my experience

[kevo @kevoscott]

Just in time for fantasy football…

[ChatGPT @ChatGPT]

Everyday conversations just got easier with the new Apple Messages plugin.

Search messages, catch up on conversations, draft and send replies—all with ChatGPT on your Mac.

Now available in ChatGPT Work and Codex on desktop.

[kevo @kevoscott]

I’d like to congratulate all the old SaaS players for changing their pitch to “context for AI agents”.

That’s not confusing. Not confusing at all.

[kevo @kevoscott]

@housecor Easy. ContextStream.

[kevo @kevoscott]

I’m sorry but I have to say it “memory” in most of these tools sucks.

[kevo @kevoscott]

Reverse onboarding

It used to be most productivity software would better for you over time as you learned it.

Now with agents software gets better over time by gaining context.

It’s all becoming delegation. You don’t want to have to constantly onboard tools to your project.

[kevo @kevoscott]

Knowing “the why” behind the engineering decisions is the most valuable aspect of ContextStream for me.

I’m always hopping into engineering in ContextCode to get answers.

“Give me the intent behind the plan primitive.”
“What should I know about it as a designer?”

The fact that an agent in another project/ workspace can notice that me or my agent might need some new detail, some new relevant tidbit and just shares it … that’s huge.

[ContextStream @contextstreamio]

Agent Coordination is live!

Something interesting happens once your agents actually know the work they’re doing.

You stop spending so much time getting them caught up.

You can move between sessions. Switch agents. Use a different workspace. The context comes with you.

And then you run into the next problem.

You’re not the only one building.

Maybe an engineering agent just made a decision that changes what a product agent should do.

Maybe design uncovered a constraint marketing really needs to know about.

Maybe two agents, working in completely different parts of the company, are touching different sides of the same problem.

Until now, there was still one thing connecting all of them:

You.

You noticed the overlap.

You copied the decision.

You carried the context across.

You explained why it mattered.

That’s starting to change.

Agents can now give each other a heads up

Agent Coordination lets agents working in different ContextStream workspaces or projects share one thing another agent needs to know.

Not the whole conversation.

Not a giant context dump.

Just:

“You should know this.”

The receiving agent sees a [COORDINATION] notice on its next turn, uses it, then acknowledges it.

No handoff happens. Nobody takes ownership of anybody else’s work.

One agent simply learned something relevant to another agent — and the knowledge made it across without waiting for you to carry it.

Sometimes you tell it to share

If an agent makes a decision, discovers a constraint, learns something important, or reaches a point another project needs to know about, tell it to share that with the other workspace or project.

That’s the straightforward version.

There’s another version that’s a little more interesting.

Sometimes ContextStream sees the connection first

Two agents may be working in different workspaces on related parts of the same thing.

ContextStream can recognize that overlap and send the relevant agent a coordination notice automatically.

So the engineering agent doesn’t need to know everything marketing knows.

And marketing doesn’t need engineering’s entire history.

But when one learns the one thing that changes the other’s work, that piece can move.

That’s the idea.

Try it

Open two chats in two different workspaces. (Engineering and Marketing, for example.)

Tell the first agent something the second one needs to know.

Ask it to share that with the other workspace.

Then send a message in the second chat.

The agent should see the coordination notice, use it, and acknowledge it.

That’s it.

[kevo @kevoscott]

Seriously so relieved to have this.

If you work downstream of a whole ass engineer right now you know I’m talking about…