📖 Glossary
AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.
Summary: On the evening of October 31, the DeepSeek Harness team published a design note for Super-agent: a resident Router dispatching multiple isolated Workers, driven by chat messages, email and CI results rather than by you typing. The note ends with four problems it has not solved. The sharpest is security — if the cloud-side loop is breached, every attached device is exposed. That is why KAIHE AIBOX (铠盒AIBOX) keeps the agent loop on a machine on your own desk.
First, what this is — and is not
Not a new model, not a shipped feature. DeepSeek calls it an architecture discussion.
What it describes is the floor under a personal agent. The dominant design still treats one session as the unit of work, so running three threads at once means remembering which conversation holds which fact. DeepSeek puts a number on that: a person can usually follow three to five active sessions. Single-task ceilings come from the model; multi-task ceilings come from attention.
Five modules, and what each buys
| Module | Mechanism | What it fixes |
|---|---|---|
| Two-layer | A resident Router only dispatches and speaks; Workers stay context-isolated and do one thing | Continuity stops fighting concurrency |
| Event-driven | Chat, email, review comments and CI results become Triggers, screened by code before a model sees them | Nothing waits for you to type |
| Async ReAct | Tools become start/stop, so a background run no longer blocks new input | One command stops jamming the session |
| Compute/storage split | The loop runs stateless in the cloud, state lives in a database, the environment stays put | Sleep no longer means the agent stops |
| Watchdog | A separate session outside the main system asks whether each Worker is making real progress | Loops that go nowhere get caught |
The split is the part people skip past. It concedes something uncomfortable: a real working environment cannot be copied into a cloud sandbox. Uncommitted changes, internal dependencies, local credentials — rebuilding them is expensive and lossy. So the environment stays where it is.
Two numbers worth keeping
Roughly 300 messages, 20 of them real work. Super-agent finished at p50 12.4 minutes and p95 15.7 minutes; the single-session-plus-subagent setup stretched to p95 29.5 minutes.
The more useful one: remove the subscription pass-through and completion time barely moves, but the Router makes about 30% more decisions and read-cache tokens climb from 23.23 million to 31.77 million. Dispatch produces nothing; it only costs. Making dispatch cheaper is how the same work gets done for less money.
Four open problems DeepSeek lists itself
Whether the Router should shard, whether Workers should talk directly — undecided. How to present tasks and let people intervene once sessions are hidden — an open product question. Memory spanning months — the main limit on scaling. And security.
The fourth is stated bluntly: if the cloud Agent Loop is breached or prompt-injected, every attached device faces takeover risk. Even without a breach, credentials entering the model context can leak in normal operation. Their direction is letting the agent use credentials without touching them.
Which machine is actually missing
The design assumes cloud and device trust each other, and that trust costs someone the job of owning the device.
KAIHE AIBOX is that machine: a dedicated Agent Computer with the loop running on your desk rather than borrowed from your laptop. Credentials sit in a local vault and the agent never sees them in plaintext. Hand over work in the daytime and it runs at night, when model calls cost roughly half the peak rate. Running dispatch on your own hardware has a quieter benefit too: the overhead lands on your electricity, not on a per-token invoice.
Three checks you can run yourself
- You juggle three or more threads that each span days. A single-session design will hit the three-to-five ceiling.
- Your work needs local ground truth — uncommitted diffs, internal systems, local files. The environment will not move, so the agent needs somewhere to stand.
- You mind credentials entering a cloud loop. Would you hand an API key to a loop that can be prompt-injected?
If two of the three line up, what you are missing is not a stronger model. It is a machine that stays on and belongs to you.
Related Reading: - DeepSeek open-sourced the Harness agent framework - Models keep getting better, but an always-on base is what you lack - When an AI agent operates everything, who holds its keys?
KAIHE AIBOX #DeepSeek #AIAgents #MultiAgent #AgentComputer
For more information, search [KAIHE AIBOX] or contact: [email protected]
KAIHE AIBOX · 7x24 Personal AI Assistant | AI Agents