Disclosure: I maintain AnyModelBot. This is an unofficial project, not affiliated with DeepSeek.
Most chatbots wait for you to speak and stop when they answer. Plenty of real work is not a question and an answer: research takes half an hour, prices need watching every day, a weekly report is due on Monday. You want to say it once, go do something else, and hear back only when there is a result or a decision that is yours.
DeepSeek recently published dsh-superagent-prototype, a proactive coding agent for one GitHub repository, along with an article called "Exploring Super-agent Design: Towards a More Autonomous Agent System". The README is clear about its status: it is a research prototype for reference, not an official DeepSeek product, and parts of the article are directions rather than shipped features. It is packaged as a bundle for DeepSeek Harness, listens to GitHub and optionally Feishu, and works on tickets in separate agent sessions.
The part that stuck with us is a way of working rather than a feature: take as little human attention as possible, and give as much value as possible. An agent that can act for a whole team can also waste a whole team's attention, so it acts only when asked, or when a rule the team wrote allows it, and it speaks only when someone has to decide or the result is ready to use.
AnyModelBot is an open-source (MIT) desktop AI assistant for Windows, macOS and Linux. We built a one-person version of that idea into it, so it works as an out-of-the-box DeepSeek Super Agent. We believe it is one of the first apps to do this for individual users. Here is what that means in practice.
Three pieces, no extra setup
Background tasks. You hand over something that takes a while: research, a report, a code change, filing. The bot opens a task with its own agent session and folder (tasks/), and the result comes back to the conversation you started it from. Chat never waits for tasks. Each bot runs two background tasks at a time, four across all bots.
Rules. A rule is a standing instruction in one sentence plus a schedule: "every morning at 8, check X's price and tell me if it is under 200." A rule can start in observe mode: it runs on schedule, but files are read-only, it never messages you, and actions that need approval are refused and logged. After three normal runs, a card in the "Needs you" inbox shows what it would have sent and asks whether to switch it to execute.
Needs you. One row in the sidebar collects everything that is waiting for you: a task's question, an approval, a failure, a spent budget, a rule that failed three times. Answer in the app, tap a Feishu card or reply in WeChat; once answered anywhere, it disappears everywhere.
A task that doesn't stall overnight
A worked example. You message: "Compare A100 prices across three clouds, as a table." The bot starts a background task and replies with one line. You ask about tomorrow's weather and get an answer immediately. Midway, one provider's prices are behind a login. The bot asks: "Skip it? If you don't reply I'll skip." Then it carries on with everything that doesn't depend on the answer. If you never reply, the default applies, the result says so, and the question is withdrawn. If you do reply later, the task picks the answer up and revises.
Quiet by default
Only three kinds of message are pushed: a usable result, a decision you must make, and a failure you must handle. Progress updates, "still working" and retries are never pushed. Quiet hours default to 22:00–08:00. Each chat gets at most 3 proactive messages a day on WeChat and 5 on Feishu; extras are merged into a morning digest.
WeChat deserves honesty: it only lets a bot message you within 24 hours of your last message, up to 10 per window. We treat that as a hard rule and don't work around it. Undeliverable results are kept, shown in the inbox, and delivered in one combined message when you next write. Feishu has no such limit.
Permissions and money
Permissions only tighten: a task inherits the bot's permission level at creation, an observing rule is read-only, and dangerous operations such as recursive deletes always ask first. Default spending caps are $1.00 per task, $0.10 per rule run and $3.00 per rule per month, estimated from Modelflare's public prices (the bill is the source of truth).
What we took, and what we left out
dsh-superagent is a team tool aimed at one repository. We took the skeleton: a status table for work, restart recovery that resumes from the oldest item, a reserved slot for messages to humans, the "observe" acceptance mode, messages only for "done, needs a decision, can't do it", and a panel that answers "what needs me right now" and stays out of the way otherwise.
We did not take Source, Trigger, Thread, Eval, Mandate review, Subscription, the GitHub PR lifecycle or Policy versions. A one-person assistant listens only to its owner. Event-driven "workloads" (for example a GitHub repo's new issues) are a direction we want to explore in the open, not a promise.
How it runs
The agent runs on DeepSeek Harness. Each bot is one agent process, and tasks and rules open their own sessions inside it. Harness runs the sessions; the app owns tasks, schedules, delivery and budgets. Background sessions report through a structured task_report tool (progress, needs you, done, failed, at most one question), and the inbox reads those fields instead of guessing from free text.
Try it, and the caveats
Download the app from the GitHub Releases page and log in to Modelflare. You get 70+ models (GPT, Claude, Gemini, Grok, DeepSeek, Kimi, GLM and more) behind one account, priced per token, and you choose the international route or the one optimized for mainland China. The caveats: tasks and rules run on your computer and only while it is on, and which features work depends on your model, operating system and permissions.
- Repo: https://github.com/modelflare-ai/AnyModelBot
- Design doc (English): https://github.com/modelflare-ai/AnyModelBot/blob/master/docs/design/background-tasks-and-proactive.md
- DeepSeek's prototype: https://github.com/deepseek-ai/dsh-superagent-prototype
If you've built something similar, I'd like to hear what you cut.
Top comments (0)