Seriously speaking, I feel the market is sorely lacking an Agent for the large number of existing devices running WinXP, Win7, and the like.
Not for developing on them, but for all kinds of debugging work it would be very useful; nowadays I'm used to just directing an Agent for everything.
Couldn't agents just Remote Desktop or ssh into them?
For seriously older stuff, I'd run the disk image in a VM.
If they interact with legacy systems and/or virtualizing is no option and/or they have no internet access, a simple USB-to-COM adapter can create wonders. I remeber doing that on a win3.1 machine using TERMINAL.EXE many many years ago.
In general, I've wondered whymost (all?) harnesses perform the API calls to the provider from the same machine that runs the tools. Why not have a client/server architecture? Then one session could talk to as many tool harnesses as it wanted to and you wouldn't have this kind of arbitrary limit of the session being tied to one specific workspace.
My harness actually has 3 distinct components: the LLM gateway, the agent session manager, and the tool server, and they can be run on different machines. The most obvious use case is the LLM gateway being a central router for all provider queries with multiple machines running the agent sessions.
The agent session manager talks to the LLM gateway over HTTP using a standard internal protocol, then the gateway has adapters for each upstream provider and translates between their external API format and the standard internal one.
The tool server is a separate jar running as a separate server process with its own systemd unit and the agent session manager talks to the tool server over a beanstalkd queue. The primary motivation for having 2 processes was total separation for unprivileged agent tool call execution (ie. the tool server systemd unit runs as a user with no sudo whereas the agent session manager can execute as root, so the user can run scripts through the interface that do things like install packages and access secrets, but the agent can't do that.
The tool server could just as easily be talking to beanstalkd over a Cloudflare tunnel or ssh port forward, though, and a session could be routed to different machines, allowing execution of tool calls in different environments within the same context window/agent session.
A typical scenario is that you walk into a small company and find they're still using Windows 7 or an early version of Windows 10. Then you have to help them deal with all sorts of problems.
I've just tried to get it to run on Windows XP SP3. (I have a VM for an old Canon all-in-one printer.) I grabbed a legacy Tcl 8.6.6 installer from https://github.com/dbohdan/tcl86-nsis-installer (mine) and a 2017 build of curl with certificates from https://winampplugins.co.uk/curl/. The harness was developed against curl 8 and didn't work with 7.53.1. MiMo-V2.6-Flash, running in the same harness on modern Linux with Wine for `curl.exe`, found that the necessary change was minimal and only touched `--write-out`:
My Free Pascal / Delphi agent compiles to a huge amount of platforms and no NodeJS required. I haven't tried targeting older platforms specifically but I do run it on Windows 10 and Windows Server 2019.
I've been working on a coding agent called Psi[0] (port of Pi) that is written entirely in ANSI C and Lua, MIT licensed. Tested on Linux, macOS, Windows, Haiku, BSD. The first three being the same executable too running via cosmopolitan libc. Otherwise it just requires a POSIX environment.
Almost all agents require Windows 10 1809 or later. Debugging issues on older devices often turns out to be a real headache, and the only option is to get SSH working.
There are Golang ports for WXP and up (check the Reticulum-go GH repo and the users), and there's MIn-C as a port of core OBSD tools for XP (Unix tools, compiler, vi, openssh and a bunch more).
For old Windows, "setting up an SSH tunnel" itself takes quite a bit of fiddling. Especially when the device you need to deal with is on a distant external network, and you can't even remote in—you're limited to video calls and the like.
I'd like a freepascal, zig or odin coding agent that can run natively on a pentium3 with 1G ram natively on either cpu or gpu - on dos, freebsd or linux at > 10 t/s.
can this conceivably be done and how?
Gemini says it can be done.
I also think the ai compression algorithms focus on PPM, BWT or CM as these take less cpu and memory then Tr and LSTM
https://mattmahoney.net/dc/text.html
Solo-c running under an Atom netbook with 1.5GB of RAM (zram). Just use bubblewrap to sandbox it but better if you run it under another user under a TTY with no $DISPLAY exported at all.
It feels very cute and mischievous to start agents in unusual environments (DOS, Linux running on a feature phone, Windows 95, etc) and see how they react :)
Yep, I've made one in win32 earlier this month for that exact reason. It's always amusing seeing DeepSeek's CoT realise it's running inside NT 4.0 once it does a directory listing. It handles it pretty well too, I gave it a intentionally broad prompt to "write a 3d shooter game" pretty much, and it went and found the DirectX SDK that comes with VS6 and oneshotted it.
this is amazing! my mind is blown by all the things that can come from this... and with this video resolution, Claude can practically read your screen in real time, crazy!
Very cool. It would be interesting to see how far it can take demo scene type projects that stretch the hardware. See this project and prompt for ideas: https://github.com/Noniv/abyssal-garden. But the harness is competing for the same resources, so perhaps better from a distance/ to a vm?
Man the first thing that came to my mind was. From this stage if we ask AI to self build to current windows OS or Mac level. I can't help but put this fake scenario in my mind that imagine if we had this back in the day.
For seriously older stuff, I'd run the disk image in a VM.
If they interact with legacy systems and/or virtualizing is no option and/or they have no internet access, a simple USB-to-COM adapter can create wonders. I remeber doing that on a win3.1 machine using TERMINAL.EXE many many years ago.
The agent session manager talks to the LLM gateway over HTTP using a standard internal protocol, then the gateway has adapters for each upstream provider and translates between their external API format and the standard internal one.
The tool server is a separate jar running as a separate server process with its own systemd unit and the agent session manager talks to the tool server over a beanstalkd queue. The primary motivation for having 2 processes was total separation for unprivileged agent tool call execution (ie. the tool server systemd unit runs as a user with no sudo whereas the agent session manager can execute as root, so the user can run scripts through the interface that do things like install packages and access secrets, but the agent can't do that.
The tool server could just as easily be talking to beanstalkd over a Cloudflare tunnel or ssh port forward, though, and a session could be routed to different machines, allowing execution of tool calls in different environments within the same context window/agent session.
https://github.com/kristopolous/old-agent
I've used it on plenty of old things.
I'd probably do a pure .vbs version if it was windows or a static pure win32 exe.
I took the same path with my for-fun/learning GUI coding agent in Tcl/Tk: https://wiki.tcl-lang.org/page/Featherweight. It shells out to curl for the networking.
I've just tried to get it to run on Windows XP SP3. (I have a VM for an old Canon all-in-one printer.) I grabbed a legacy Tcl 8.6.6 installer from https://github.com/dbohdan/tcl86-nsis-installer (mine) and a 2017 build of curl with certificates from https://winampplugins.co.uk/curl/. The harness was developed against curl 8 and didn't work with 7.53.1. MiMo-V2.6-Flash, running in the same harness on modern Linux with Wine for `curl.exe`, found that the necessary change was minimal and only touched `--write-out`:
Here are two screenshots of the result. Featherweight is running in Tcl 8.6.6 on Windows XP.https://paste.dbohdan.com/1ng3b5y-kb4yd/greenshot-1.png https://paste.dbohdan.com/1ng3b5y-kb4yd/greenshot-2.png
https://github.com/FMXExpress/PasClaw/
[0] https://psicode.ai/
It would not make a whole lot of sense to me.
I also think the ai compression algorithms focus on PPM, BWT or CM as these take less cpu and memory then Tr and LSTM https://mattmahoney.net/dc/text.html
New hotness: Running an AI agent on every piece of old/tiny gear.
Finally it can outperform TempleOS.
Someone should rewrite the old Lemmings game with AI as well.