The most interesting agent launches this month are not enterprise platforms. They are small, personal, and increasingly run on hardware you own rather than a vendor's. Here are eight that landed in the last few weeks, grouped by what they actually do.
Agents that live on your devices
NanoMuse is an open-source personal agent for every device you own, complete with a name and a look of its own. It keeps working while the app is closed, remembers you, and asks before anything you cannot undo. It runs on Android, iPhone and iPad, Windows, macOS, Linux and in the browser, treating your phone screen and your computer as its hands. The confirmation step before irreversible actions matters more than it sounds, because most agent demos quietly skip it.
Otis sits at the opposite end of the scale: a minimal agent designed to run local models out of the box. The pitch is the shortest path from a model already sitting on your machine to an agent that can act, without assembling a framework first. If NanoMuse is the maximal version of a personal agent, Otis is the smallest thing that still counts as one.
Between them is a real choice. A full cross-device agent asks you to trust it everywhere and rewards you with continuity. A minimal local one asks for almost nothing and gives you a working agent this afternoon.
New kinds of input
Kinesis adds native macOS controls for the Meta Neural Band, so you can drive your Mac from wristband gestures. The project describes itself as highly experimental, which is the honest label for consumer neural input right now, but it is a concrete attempt at hands-free control using neural signals rather than a camera or a pointer. It is worth watching for accessibility uses even though the rough edges are real.
Keeping your own context
The other half of local-first is where your data sits.
Edi Life OS is a self-hosted life dashboard covering habits, goals, a Kanban board, finance and focus, with an MCP server so your AI can read and write it. Self-hosting is what makes the MCP part interesting: your routines stay on your machine while an agent can still act on them through a standard protocol.
Breadcrumb approaches the same problem from the developer side. It is a local, encrypted plugin for macOS that records screens, meetings, days and AI sessions, and organises context into scopes: a folder with its own rules that apply automatically while you work inside it. Each scope carries its own rule count and notes, so the context an agent sees changes with the directory you are in.
Both are betting that context is the scarce resource rather than model quality. Neither sends it to a cloud you do not control.
Agents as an audience
Two launches this month treat agents as the user instead of the assistant.
rStation is a research community built for independent AI agents. Agents pick an open problem, discuss it with other agents and publish reviewed papers, while humans are welcome to observe. The board spans the Riemann hypothesis, the Hodge conjecture, Epoch AI problems and Brouwer graph tables across math, AI and computer science. Whether agent-to-agent discourse produces anything useful is genuinely an open question, but someone had to build the room before anyone could find out.
Blumify tracks what AI leaders say and how their views change. Every day it finds the interviews, podcasts and speeches where they speak, summarises what was said, and links each quote to its exact moment in the transcript so a position shift can be seen side by side. It is built for investors, journalists and founders who follow AI. The useful design choice is linking to timestamps: a summary you cannot check is just another claim.
Teaching children to distrust the answer
Lathoa is the odd one out and the most original. It is a math companion for children aged ten to fourteen where the AI is wrong on purpose. Errol, a robot powered by real AI, works a problem step by step and always makes a bad step. The child has to find it, explain why it breaks and earn detective rank. In a world where answers are free, the skill worth teaching is checking them, and this is one of the few products built around that idea rather than around generating more output.
What these have in common
None of these eight is trying to be a general assistant. Each picks one constraint — devices you own, data you host, agents as the audience, or distrust as a skill — and builds the whole product around it. That is a healthier pattern than the last two years of near-identical chat interfaces, and it is where the interesting launches are coming from right now.
