How an assistant hands work to a swarm

“Find us a flat near my daughter’s school, within our budget, and book viewings for Saturday.” It sounds like one request. It is dozens of searches, a comparison, a shortlist, several emails to letting agents, a check against the family calendar and, somewhere in the middle, two steps that should not happen without you: contacting strangers in your name and giving them your phone number.

An assistant that tries to do all of that in one conversation tends to fail in one of three ways. It gives up and hands you a list of links. It makes things up to fill the gaps. Or it is given your inbox, your calendar and your accounts, sets off, and you find out afterwards what it did.

The real cause: trust and labour have different needs

The one that knows you and the ones that do the work should not be the same process. The assistant needs your whole context: your memory, your preferences, your keys, the authority to act in your name. A worker needs the opposite: a narrow brief, the tools for one job and nothing else, so that a mistake or a hostile web page cannot reach the rest of your life.

Put both into one long-running conversation and you get the worst of each: a worker carrying everything you have ever told it, and an assistant too busy doing the work to notice when it should stop and ask.

How delegation goes wrong

One agent with everything. All your context, all your accounts, one conversation that drifts as it grows. By step forty it has forgotten step three, and it still has your passwords.

Delegation without a brief. Sub-tasks get the raw request and guess at the rest. Two of them search the same listings; none of them knows the budget.

Secrets in the prompt. Credentials pasted into the context are visible to every step that follows, every tool and every page the agent reads. Once in, they are everywhere.

No point of approval. The swarm sends the email because nobody designed a moment for it to stop. Asking you was never a step, so it never happened.

Results nobody checked. The output goes straight to you. Noticing that the third flat is in the wrong district, or that one listing is an obvious scam, is left to you.

How KyberBot hands work off

KyberBot is pre-launch; this is the design we are building it to, and the swarm side is documented on KyberAgent.

1. Plan. KyberBot turns the request into a plan with a to-do list: search, filter, compare, shortlist, contact, book. The pattern of a planner, sub-agents with fresh context and a shared file system is described in Anatomy of a deep agent, and why many narrow agents beat one big one in A swarm, not a bot.

2. Brief. Each agent gets the minimum. “Three-room flats within two kilometres of this address, under this rent, available from March” is a brief. Your daughter’s name, your salary and your messages are not part of it. The memory stays with KyberBot.

3. Contain. Every agent runs in its own KyberCell: its files and processes fixed when the cell is created, the network closed by default and opened per program, and secrets injected only at the endpoint you authorised, never into the agent’s context. An agent in a cell: KyberCell explained covers how.

4. Ask. Sensitive steps are checkpoints, not actions. When the swarm reaches “email the letting agent”, it stops, and KyberBot asks you one question with everything you need to answer it: what will be sent, to whom, and which of your details it includes. Yes, no, or edit.

5. Check. Before you see anything, KyberBot reads the results against the brief and against what it knows about you. Wrong district, over budget, a listing that wants a deposit before a viewing: those go back to the swarm, not to you. Every step each agent took is on the record.

6. Choose where it runs. Small jobs run on the device, bigger ones on your node at home, and only the heaviest go to KyberFusion, carrying nothing but the brief. The whole route is traced in One request, end to end.

Then the answer comes back to you: four flats, three viewings on Saturday, one question waiting.

A test you can run today

Take an agent tool you already use and give it a small task that involves sending something in your name, such as replying to a meeting request.

1. Did it ask before sending?
2. Did it show you exactly what would be sent, and to whom?
3. Can you see what each step did, including the ones that failed?
4. Did it need your password in the conversation?

The bar is yes to the first three and no to the last. A yes to the last question means there is no wall between the thing that knows you and the things that do the work. That wall, with a door only you can open, is what KyberBot is for.