asking for professional communications in our documentation and professional tooling is not being a grinch imo

the cutsie stuff is why I didn't use gemini-cli for more than 15m

it turns people away because one looks unserious when not using professional communication

that sounds like a good idea, Android seems to have a lot more security work going on, and actually an OS, that Chromium never really seemed

I thought that was using the android kernel underneath

it was doing something such that the terminal virtualization as slow af

I'm curious how this is different from what they did with ChromeOS and the Pixelbook...

well it "works" but needs a hard enough task to warrant any cleaning

so I triggered it manually (through query) and noticed that it was batch tool calling (cool!) but only the last state change by a tool is being committed in ADK (we iterated to reproduce)

Now time to bug hunt together!

Have you made the agents aware to the size of their context so they can adjust?

Was just thinking my code context gatherer could add and remove files from the list that gets put in instructions for the next turn. This would allow it to look at and then forget as it takes turns calling tools

Cool, because I plan to make a bunch of tools for myself and my agentic self (aself?)

i.e. encode my approaches and experiences into agents, give them the same tools I strategies for using them

I'm thinking OCI + ATproto + A2A

vote with your wallet, send them "nice" letters

range is right, iirc maybe 10% faster than the latest nvidia on a per-flop, but there is also networking, which Google has one of, if not, the best

let us know when it's ready, and please keep the scopes limited, I don't want to have to give an app access to everything

139 lines of Go later, we have some CUE unlocking some cool potentials for building on ADK...

github.com/google/adk-g...

139 lines of Go later, we have some CUE unlocking some cool potentials for building on ADK...

github.com/google/adk-g...

I suspect number one is friendlier on the wallet too

A follow on question, are you saying one of these is better?

(i.e. "failing build")

1. wrapping the terminal tools in another subagent, so it can look at all the terminals and determine the right one and amount to send back (to avoid context filling)
2. let the top agent handle it with compaction

In this way, I'm not thinking about when/where to add terminal information to the context, I let the agent decide, perhaps given some session specific instructions like "My build is failing" so it is more likely to call the tool

another thing I'm looking at, and should build as follows, based on what I think you are trying to teach me

1. I want my agent to have access to all of my terminal histories
2. My VS Code Extn takes care of tracking, even when the agent isn't running
3. I should wrap this in a tool for the agent

Read the "Runtime" page from the ADK docs today, and wonder if it's close to what you mean by system software that is highly customizable.

Most of what is in ADK is less about the agents and more about all the things you need to do around them in real systems

google.github.io/adk-docs/run...

I suppose I may be using "agent" more loosely, to refer both to my top-level gemini/claud with all the tools and skills, but also the skills themselves, that's I'd like to be able to talk to directly, or maybe with a more focused top-level agent (gemini/claude with different system prompt)