Agents have been hot for almost two years, but when people actually try to run one, many stall at step one. What to install. How to make it really do work instead of just chatting in a box.
How to pick among four
In recent months the open-source world shipped several agent runtimes with different ideas. I picked four representative ones to help you choose by your own situation.
LOOP: the Rust workhorse
First, LOOP. Made by Indian firm Soket AI, written in Rust, released in early October. Its most appealing trait is resource thrift. Officially, under the same model, it uses only one-seventh-point-seven the memory of Claude Code and one-seventeenth the CPU. Long tasks can run for hours, days, even weeks, with Git-style branching and resume, and multiple agents can split work as a team. Who is it for. You have a modest machine, want local long runs, and fear blowing up memory. It is the top pick.
Pi: minimal local
Second, Pi. Earendil's 1.0 stable from October 1, open source under MIT. It pushes minimalism, letting the framework fit your workflow rather than the reverse. A Codemode feature lets the model write JavaScript to call tools, and the team says prompt tokens drop about forty percent, from over five thousand to around three thousand per request. It supports more than fifteen model providers. Who is it for. You want a light, self-controlled harness not locked to one vendor.
DeepSeek and Agent-Native
Third, DeepSeek Harness. Open sourced August 13 under MIT, written in TypeScript on the Cordis architecture. Its boldest slogan is everything is a plugin, models, tools, skills, and UI all add or remove as plugins. One command, npx @deepseek-ai/dsh web, pulls up the interface. Its star count exploded past two hundred thousand. Who is it for. You want to tinker deeply, build your own plugins, and do not mind APIs shifting during public preview.
Agent-Native
Fourth, Agent-Native. Builder.io's open-source beta from October 14, also TypeScript. Its angle is different: the agent and the UI share the same actions, data, and state. You change something in the interface, the agent sees it at once; the agent finishes work, the result shows up in the app. Who is it for. You are a developer wanting to build agent-powered apps a human operates alongside.
A blunt verdict
How to choose, a blunt verdict.
Mind permissions before you start
One caveat for all. These are public previews or early versions, interfaces will shift, do not migrate your real production flow onto them in one move. Test on a non-critical small task first, in an isolated directory, with recoverable data before loosening up.
Permissions, seriously
Another reality, permissions. A framework that runs scripts and bulk-edits files is only as safe as the file scope you grant it. Start small in trials, do not hand it the whole disk on day one.
The easiest on-ramp
If you have never touched such a framework, here is the laziest on-ramp. Skip the highest star count. Start with Pi or LOOP, install in one command, run a tiny task like sorting desktop folders or batch-renaming files, and feel how an agent actually moves.Go deeper only once comfortable. Want your own plugins and models, then look at DeepSeek Harness. Want interface-bearing apps, then touch Agent-Native. Add layers, do not try to build the system on day one.A cost people miss. The framework is free, but the model it calls costs money. An hour-long task keeps burning tokens. Test locally with a small model or a capped plan first, learn the burn, then scale.Finally, all of these run on local models. LOOP and Pi both support Ollama style local inference. If your machine is strong enough, sensitive data stays on device, and privacy and cost both get steadier.
The language ceiling
One overlooked pick factor is the language ceiling. LOOP is Rust, to patch the core you need Rust. Pi, DeepSeek Harness, Agent-Native are TypeScript, easier if you come from frontend. Choose by your own stack, not just features.Community health is worth a glance. High stars are attention, what helps is whether issues get answered, docs are complete, examples are many. Spend two minutes on the repo before download, beats stepping on traps after.Also, do not get drunk on everything is a plugin. Many plugins does not mean you need them. Know the concrete problem first, then see which framework's defaults sit closest, good enough is fine, greed makes maintenance hard.
Start small
When you start, do not plan a big system up front. Take a weekend small job, like letting an agent tidy your downloads or rename a batch of files by project. Run that, then think bigger.Keep a notebook during it. Which framework installed smooth, which docs confused you, which step stuck longest. More useful than ten reviews, because you record your own feel.
Ten minutes in the repo
Before picking, spend ten minutes in its repo. Examples, docs, whether questions get answered. That step saves the half day of wasted tinkering later.
Run one small task first
Run one small task first, then talk big plans. How fast onboarding is, you learn by trying, not by reading star counts. Your own machine, your own job, feel beats leaderboards. Skim the repo before install, if docs confuse you switch early, the saved tinkering time buys two real tasks.
Do not greedy-pick
Do not greedy-pick frameworks, know one concrete job you want done, then match which fits closest, stop when good enough. Many stall in endless comparison and finish zero real work, one small run beats a week of reviews. After real use, note the pitfalls, which installed smooth, which docs were poor, next time no re-stepping, experience is yours and worth more than stars.