An agent is a key with a job. Revoke the key, the agent is gone.
Agents are only as safe as the memory they run on. Agentu agents work on encrypted memory through scoped keys: they see what they are allowed, decide within the limits you set, ask you for the rest, and can be removed — with everything they knew — in one revocation.
For your life: an agent that is literally your memory
Not a chatbot that forgets on Monday. An agent that knows your prior decisions, your preferences and your limits, and runs the parts of life you hand it.
- Decides the way you would. It reads your encrypted memory — how you decided last time, what you always say no to, which bills get paid on which day — and acts on those parameters.
- Asks when it should. Anything outside its limits becomes a yes/no on your phone: "Renew the insurance at €412?" — and the answer is remembered for next year.
- Pays bills, sorts admin, chases the things you keep meaning to do. Within spending limits and merchants you approved, with every action in the signed history.
- Runs on any model, keeps nothing with the model. It calls the model you chose, under business terms that forbid training, and all state lives in your encrypted memory — never in the provider's chat history.
- Yours to switch off. Pause it, narrow its keys, or revoke them. There is no copy to find.
How a personal agent is built on Agentu
- You create keys for the categories it may read and write —
read:Financial,write:Admin— and nothing else. - You set limits: amounts, merchants, hours, "always ask" lists.
- The agent runs on the model you choose, through Agentu, and records every decision and every approval in the memory itself.
- Your phone gets the questions; your memory gets the answers.
The memory, keys and model connections exist today; the agent runtime and the phone approvals are planned. Payment actions will launch only with regulated payment partners, in the countries where they are licensed.
What a business agent team looks like
- Research agent —
read:Market, Perplexity for the open web, Claude for synthesis. - Drafting agent —
read:Matter-2041,write:Drafts, the model your lawyers rate best. - Finance agent —
read:Ledger-2026, a cheaper model for reconciliation, a stronger one for anomalies. - Support agent —
read:Product,write:Tickets, anonymises customer details before any model sees them.
Each agent is a key with a job description, a budget and a model routing policy. The audit log shows what each one did; the memory shows what each one learned.
For business: entire agent work teams, on your encrypted data
Build teams of agents that use all the leading models — the best one for each job at the best price — while your data stays encrypted, anonymised where it must be, and scoped so no agent, model or person holds more than its slice.
- Best model, best price, per task. Routing policies send each job to the model that does it best for the least, through one account and one set of privacy terms.
- Anonymised by default. Names, identifiers and numbers can be replaced before a model sees a prompt and restored when the answer comes back.
- Encrypted by department, level, role and project. Agents hold keys exactly like people do; so do the models they call.
- Removed in a second. A contractor's agent, a finished project's team, a model you stop trusting: revoke, and it is over — in every jurisdiction, under every law that requires it.
What exists today
| Capability | Status |
|---|---|
| Encrypted memory with scoped keys per model, person or agent; provenance; signed history; instant revocation | Live |
| Agents built by you on the API, SDK or MCP tools, with their own keys | Live |
| Team, role and project structure for company memory (Agentu Professional) | In build |
| Personal agent runtime with phone approvals and limits | Planned |
| Business agent teams with model routing, budgets and anonymisation | Planned |
| Payments through regulated partners | Planned |