Why it mattersA team that runs many agents at once can now check their work in one place. The engine keeps a record of who did what and who approved it.
Most AI coding tools give one developer one chat with one agent. SoduWork, a new open-source project from MDiRi AI, is built for teams that want to run many agents at once. It routes each task to one of nine agents. A second agent reviews the first's output. The request closes only when a named person signs off.
The repository went public on 2026-10-29. In sixteen days it has reached 773 stars and 124 forks on GitHub, under an MIT licence. The project is in public beta. The README says "SoduWork is in public beta. Try it and evaluate it before using in production."
What the engine does
SoduWork sets up each request in three parts. A person or an agent writes acceptance criteria before the work starts. One agent then does the work. A second, different agent reviews it against the criteria. The request closes when the person who asked signs off. The README puts the idea plainly: "A chat ends with an answer. A SoduWork request ends when the person who asked signs off."
A method that works once can be saved as a plugin, with a review step before the team shares it. Each action records who did it. Admins can set what an agent is allowed to touch.
The nine agents it can drive
MDiRi AI calls each agent runtime a harness. The nine on the list are Pi (its own built-in harness), OpenCode, DeepSeek Harness, Qwen Code and Kimi Code. The last four are Codex Harness, Codex, Claude Code and Google Harness². The README says they share one adapter interface. A user can switch from one to another mid-conversation and keep the history.
What's running underneath
SoduWork runs from source on Python 3.12 or newer, Node.js 22.19 or newer and Git. It is published as a desktop app for macOS, Windows and Linux. It also runs under Docker for teams. A live demo is on soduwork.ai. A cloud runner for long-running tasks is in the project. The project has 523 commits.
The code is MIT. The name and the logo are MDiRi's trademarks and are not part of that licence. There is no price on the public page, because the user supplies the agent accounts and the hardware.
What the README does not say
MDiRi reports no named customer and no production user count. There is no benchmark against Omnara, Claude Managed Agents or any other multi-agent engine. There is no failure rate for the review step. The public beta note is a warning that the tool is not yet proven in production use. The README leaves it to the reader to decide whether to trust it.
Why this matters for a working team
A team running several agents today usually ends up with one chat window per agent and no shared record. SoduWork pulls them into one place. The split between the doer and the reviewer is the piece most team-scale tools do not have. The first agent writes the code. The second agent checks it against the rules the team set. A person approves at the end. The record of each step is kept.
The harness list matters because teams rarely standardise on one agent. One person on Claude Code, one on Codex, one on DeepSeek Harness is a real setup. Routing all three through one engine lets the team share one set of acceptance criteria. The record of who approved what is in one place too. The cost is setup time and the extra review step. The gain is that the team can tell which agent produced which change.
Source
mdiri-ai/SoduWork on GitHub. The project's live demo is at soduwork.ai.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.