https://github.com/smol-env/smol
here are traces from an agentic task around using duckduckdb
comparing CPU and RAM usage of the whole container over time w/ OpenCode, hermes, pi, codex, smol
https://x.com/__tosh/status/2086882367126286466
https://x.com/__tosh/status/2086882204060160350
smol is very minimal only using stdlib (in this case it is the go version but you can also take a look at implementations in python, clojure, php)
(any OpenAI Responses API compatible endpoint works, if your endpoint does not support 'custom' tools you can have your agent change the smol implementation to use 'function' tool implementation instead)
that said: be aware that smol does not come with any system prompt and does not load agents.md files by default
some older not so strong models benefit from a system prompt and guidance in agents.md that complements them
that said 2: system prompt and or loading agents.md automatically is easy to add though if you want it
This is a coding agent implementation I am working on, which delivers what this promises (at least on the "lean" part), except it's actually fully open source, and even more lean (few MBs of runtime memory usage).
MIT-licensed, written in C, multi-provider / multi-model, minimalist approach to system prompt and tools (think kinda like pi, but with a bit more "batteries included", like subagents and background tasks out of the box), polished presentation, inspectable (usable transcript view), etc.
Also, security aside, engineers want the freedom to modify and experiment with the tools we rely on.
Tools like this are too important to be closed. Do you want to be Internet Explorer or Firefox?
I can’t even guess what this means
But from a fellow engineer's perspective we don't need another commercial solution in this space. Every major tech vendor is working on harnesses, and we will get ones that will run circles around yours for free. The open source ones that get popular will have mit/freebsd/etc permissive licenses. (or no one will adopt them) But coming at it from the angle you are indicates you haven't reasoned well about this. Please don't hurt yourself and those around you by starting a startup on this idea.
Yes the core part is a simple loop, but we have all built toy compilers, inference engine, browsers (it is just a curl command eth) etc. The core algorithm is supposed to be simple.
Had a discussion recently: https://x.com/NoCommas/status/2086568454434537710?s=20
I wonder if this is a viable approach; after all frontier model providers are betting on the opposite.
Then again, they bundle their harness and offer subsidiary pricing - so maybe they themselves aren’t sure if models are as important.
Had a discussion recently: https://x.com/NoCommas/status/2086568454434537710?s=20
They can one shot entire games, with relatively minor issues.
And obviously asking for small code snippets and integrating them yourself has been well supported for five years.
But in Agent mode... not so much. I was asking frontier models to make simple changes to my Pong game (you know like the one from 1972) and it constantly failed to make simple changes or would break something else in the process.
The main issue is that they can't see what they're doing. Actually one of the agents tried playing the pong game by screenshotting every frame, and it ran for about 20 minutes before I realized what it was doing, and told it to calm down.
It takes about 10 seconds to process an image, so it was running the game at 0.1 frames per second... 600x slower than realtime. The technology is not quite there yet.
If your game is something turn-based though, with discrete States and well-defined transitions between them, they can help out a lot more with that.
Opt-out every time is unacceptable. If you opt out once it should be enough. One accidental execution path without the right environment and you're spewing telemetry to spaghetti knows where.
That would also be true with opt-in by environment variable, though.
as for why stop here yep, the goal is to be able to find the best sweet spot in being self contained v.s. all-in-one bloat ware. for example, we still use a bundled `tmux` skill for the orchestration.
Harness is compute and Model is data
Fuck you. Will there ever be a decent agent where the answer to this is "No"?
- Ante installs a pinned, checksum-verified official llama.cpp build matched to your machine (Metal on Apple silicon; CUDA, Vulkan, or CPU on Linux) and handles upgrades when the pin changes. - It discovers GGUF files already on disk (~/.ante/models, the llama.cpp and Hugging Face caches), attaches to llama servers already running on local ports, and estimates RAM/VRAM from model size and context window before anything loads. - `ante --offline-model /path/to/model.gguf "prompt"` boots the server, runs the session, and shuts it down. `/offline-mode` does the same interactively; `ante serve --offline-model` loads a model once for many clients. - No API key, no account. Once the model is on disk, inference needs no network at all; set ANTE_TELEMETRY=off and no telemetry is exported either.
On capability, we'd rather publish the number than oversell: we benchmark local models with the same harness and auditable runs as frontier ones, and Qwen3.6 27B (a 17 GB download) scores 56.2% on Terminal-Bench 2.1 across 445 trials (live results: https://antigma.ai/eval). That's a real gap from frontier models. The design bet is that you mix: hosted providers and local live in the same catalog, `/providers` switches mid-session, so sensitive repos or high-volume work go local and hard problems go frontier.
Hosted models work with your own keys or subscription. But nothing about trying Ante requires signing up for anything: download the binary, point it at a GGUF.
Offline mode is under active development and has rough edges with the overview at https://ante.run/local/overview. I'll be in the comments.
> Ante installs a pinned, checksum-verified official llama.cpp
But in README: > Ante ships its own inference engine
May I suggest you use the first phrasing in both places. I took it as Ante devs had written their own engine and I doubt I'm the only one.