๐ Mastra agents now support the Agent Client Protocol (ACP).
Run Claude Agent, Codex CLI, Cursor, Gemini CLI, or any other compatible harness directly from your Mastra agents.
Discussions about Memory seem to be focusing on mutable representations the world in a format that feels less deterministic than I would personally prefer. Might it come handy? Sure
Ideally, I'd like a few solid things to go along with my stochastic genius parrot though
Do any of the next gen ADEs offer tools or extension to prepare and manage agent contexts?
The best idea I've seen so far come from the pi-ai tree exploration + fork/clone
Would seem important to analyze, explore and design contexts? And set your agents up for success.
On the one hand, session analysis tools like this can help you understand what's going on in your sessions and would probably be best analyzed passively by agents
This isn't so much about memory - it's about how much of your world were you able to efficiently present to models
We'll see a lot more of this - A2A and U2A in any app is table stakes at this point
It's not just model routing anymore, it's harness routing too. Any harness to any harness.
Seems entirely obvious. Like giving agents access to browsers, then computers
@signulll i would bet power users will develop shorthand with their agents kind of how people used text expansion (omw -> On my way!) but for agents
i would also bet the screen recording + pointing + narrating workflow becomes more powerful as multimodal models make it cheap and efficient
@signulll the more capable the computer use and agentic JiT coding agents become, the less important the precise action explanations become and the more valuable the velocity of intention -> agent turns
imo next step virtual computer use becomes a thing (codex uses A11y to drive action)
subtweeting an entire industry exponentially adding agentic engineers to their projects ๐คฃ
bien jouรฉ
the inevitable "Mythical Agent Month" article on X going to get wild engagement
llm + context + tools is can be closer to serverless than IDE if you change your perspective
how we make this work from a tokenomics and determinism + quality perspective will require interesting engineering
hugely valuable but we'll need lots of very capable cheap tokens
things are moving in interesting directions with the new agentic stack - it reminds me of the early days of cloud when we were all just figuring out how to make python and rails work on ephemeral machines that didn't have stable anything
this is similar work but for stochastic
feels like a lot of us are eager to build composable applications that build on the harness directly without having to split code paths for hosted application vs harness
cursor's https://t.co/MRfeC19O16 and https://t.co/87GJymFFdJ are exciting steps in this direction
I'm personally very pro AI and i think the jump in quality of life and cognitive potential for humanity is going to be extraordinary *AND* providing thoughtful access to this technology is primordial for a best case scenario in human flourishing
It's unlikely that esthetics of datacenters, which are built in the middle of industrial areas, have anything to do with the citizen pushback on AI
the general argument imo is - high utility consumption + potentially pollution neighbor that doesn't really create many jobs
Many people do not seem to want data centres built near them, despite the fact that they don't cause that much traffic and often generate a lot of local tax revenue. I suspect it's partly because they're ugly! My proposal:
An argument could be made that a combination of local benefits could make this an attractive proposition
- compute take rate for community
- subsidized compute
- mixed use for education and community
given the scale of the infra being built, could build insane library network
one of the downsides of respecting privacy as much as I do is that my in app feedback submissions provide such little personal information that I don't know how to contact users in response to their feedback
if you have an M3 Max w/ 48GB and reached out overnight, please DM me
it all started from wanting to let my agents talk to each other, directly on my machine, directly across harnesses, models without getting bogged down in topology heavy frameworks that mandate things be run a certain way
that's turned into a bit more than slack for agents imo
Been working hard on https://t.co/0WL8Uivn8P and it's turning into a joyful to use and beautiful piece of software.
Getting it to work across agent harnesses w/ different modalities has been a fascinating dive into the simplicity and complexity of agent to agent communications