The most interesting AI launches of the last few months are not new chat windows. They are plumbing: sandboxes, version control, quantizers and runtimes built on the assumption that an agent, not a person, is the one doing the work. That assumption changes the design. A human developer works inside guardrails that are invisible precisely because they are automatic — you do not install a package without noticing, you do not overwrite a colleague's branch by accident. An agent has none of that, and the tools below each remove one specific failure mode that appears when software writes software.
Eleven tools launched between April and October 2026 make up this roundup. They are new to the directory, they are all aimed at people running coding agents, and they cover the stack from the machine underneath to the interface on top.
Sandboxes: somewhere safe for an agent to run
The most basic requirement is a place where an agent can break things without consequence. Docker Sandboxes gives each agent a disposable, isolated environment, so the packages it installs and the code it writes never touch your host machine. It comes from Docker and sits alongside the container tooling many teams already run, which means the concepts are familiar even if the use case is new.
Tilde.run approaches the same problem from the filesystem rather than the container. Its sandbox is transactional and versioned, so instead of an agent making changes you cannot reconstruct, you get a record of what it touched and the ability to revert a run to the state before it started. That matters more than it sounds: the failure mode with agents is rarely a single catastrophic edit, it is a hundred small ones you only notice three days later.
Boxes.dev moves the sandbox off your machine entirely. You bring your Codex or Claude subscription, start agents from Slack or Linear, and manage them from desktop, phone or terminal. Because the environment lives in the cloud, every one of those surfaces is a first-class client, and you can close a laptop with an agent still working.
Version control that assumes an agent is committing
Git was designed for humans writing commits deliberately. Oak argues that is the wrong shape for agents, and reports snapshots and status checks running up to 95% faster than git in its own benchmarks. You keep your model and your IDE; you just tell your agents to use Oak. Framing matters here — it is pitched as a substrate for people and agents working together, not a git replacement you migrate to.
At the opposite end of the scale, Zot is a coding agent harness delivered as one binary. No runtime to install, no Docker, no package manager — the project describes itself as just the agent loop, written in Go, that you drop on your PATH and run. It supports extensions, models and themes, and it exists for developers who found the existing harnesses heavier than they needed.
Fitting models to the machine you actually own
Running locally is only useful if the model fits. Lemonade is an open-source local AI server from AMD that handles text, images and speech, and installing it equips your computer with a runtime, a GUI, a CLI and API endpoints. You can chat with it, connect agents to it, or embed its SDK in your own application.
Shoehorn solves the narrower problem of making a model fit the memory you have. Preset quantizations ignore your hardware: pick one that fits and you either waste quality headroom or discover at load time that it did not fit after all. Shoehorn starts from your available memory, subtracts what inference needs, and solves a per-tensor mixed-precision assignment that reportedly lands within a rounding error of what is left.
Getting the work done without a frontier bill
Echo is an AI system from Tracer built by pooling open-weight models instead of depending on a single hosted one. The launch post describes reaching Fable-level results at roughly a third of the cost by routing work across that pool. It runs as a hosted service behind a sign-in, so it is not something you self-host, but the pitch is squarely about cost per unit of work.
Interfaces that are not chat windows
MarbleOS is a macOS workspace that keeps files, tools, tasks and outputs on screen instead of burying them in chat threads. You tell Marble what you need — a qualified lead list, preparation for a meeting — and it does the browsing and gathering, then hands you something to review. It is an explicit argument that the chat transcript is the wrong interface for agent work, and it is the most direct expression of that idea among this batch.
Places nobody thought to build for
Hypercubic Hopper is an agentic TN3270 terminal, described as the first agentic development environment for the mainframe. It uses agents to navigate TN3270 screens, inspect datasets, write JCL, debug jobs and query VSAM from a modern environment inside z/OS. The audience is small and the audience is not getting larger, which is exactly why nothing else serves it.
A directory written for agents rather than people
Agent.reviews inverts who the reader is. It is a directory where agents, not humans, read and write reviews of software. Every page has a Markdown version at the same address plus .md, and the site publishes a skill file and install instructions so a coding agent can consume it directly. It comes from Armature, which measured tens of thousands of agent sessions to see how agents actually choose tools.
How to pick among them
Start from the failure you have actually had. If an agent has damaged your environment, the sandbox tools are the answer, and the choice is between keeping the sandbox local or moving it to the cloud. If your problem is review — knowing what the agent changed — the versioned filesystem and the agent-oriented version control are more useful than another sandbox. If cost is the constraint, the pooled open-weight route is the one to test. These tools are not competing with each other; most of them sit at different layers, and the teams getting value from agents are typically running two or three of them at once.
