A continuous improvement platform for production AI systems: it distils code, warehouse, dashboards and experiments into a knowledge book with citations, then sends a team of agents to investigate in parallel and prove changes against a benchmark, handing back work that can ship.

Meta's personal agent Muse runs on a separate cloud machine with its own browser: it fills forms, compares prices and places orders, keeps working after you close the app, and asks Sentinel before anything irreversible.

An agent app from Manus for everyday life: every agent gets its own mailbox, phone number, wallet and cloud computer, so it can take calls, pay within a budget you set, and work in groups.

An open source autonomous agent from Nous Research: finished tasks harden into reusable skills that keep improving, preferences live in local files that survive sessions, and it plugs into more than twenty chat platforms.

A cloud agent from Tencent Cloud: mention it in an issue or pull request and it reads the repository and CI context, plans, edits code, opens a pull request and gets the build green, on its own.

An AI assistant platform from Tencent Cloud, MIT licensed, where conversations, workspaces and keys stay on your own machine and nine chat platforms route into the same set of experts.

An open source agent runtime from DeepSeek with a desktop build that bundles Node.js, pnpm and a Python environment, four preset modes, a plugin page and a trace view.

Search, scrape and interact go through one interface, and what comes back is clean markdown, templated JSON or a screenshot. Proxies, anti-bot handling and rendering are the service's problem — which is exactly the tedious part.

Four layers split by depth — a no-code studio, a conversational layer, an event-driven core and an extension pack. The framework is in maintenance mode, so the honest question is whether to start on it or only keep running it.

Nodes are steps, edges are transitions, state travels between them — and the same graph may mix deterministic code with model judgement. Resume points, approval gates and inspectable state are primitives, not patches.

Agents, flows and tasks form the skeleton; a control plane on top traces every model call and memory read and costs it. The framework itself is MIT, and the discovery and optimisation stages depend on platform-scale data.

Four screens split the work: describe an agent in a sentence, drag blocks for a complex flow, watch runs and spending, and reuse what others published. A platform directory under a source-available licence sits beside a classic MIT one.

An assistant that lives in the chat apps you already use and keeps every trace of its memory on your own disk. A non-profit foundation holds the project, there is no paid tier, and one gateway process routes 29 channels into a coding agent.

Give it a task and it plans, executes and hands back a finished file, because it runs inside a networked virtual computer with a disk that survives between steps. Credits replace seats, and every tier keeps the same concurrent and scheduled task limits.

One half takes a task and returns a result, the other rents your own Playwright code a remote browser with a stealth profile. The open library is MIT licensed, and the billing is per hour and per token rather than a subscription.