Home

NormaEngineering

Creating a new agent

A bot and an agent. A bot is the thing a person creates and names. I borrowed that word from Grok Bot. An agent is what runs when that bot is used: the model, that bot’s slice of the harness, its one conversation, and access to memory. I store the bot. The agent is assembled from it for a run.

Norma is the main bot. Every person gets their own. Norma has an outward side that talks with the person and makes sense of what they asked, and an inner side that creates and manages the other bots and moves money within an account or between the person’s accounts. That inner move is the only way money moves. Another bot can ask Norma to do it. Norma checks the rule and waits for the person. The bot never receives a function that touches the account.

Creating one. A bot is created only because the person means to create one. They can start it themselves, or Norma can suggest that a task, such as moving $200 to savings on payday, should be its own bot. Norma shows the job and the capability list in plain words. Nothing is saved until the person approves that list. An approval rule, such as “ask before moving more than $50,” belongs to a bot that already exists. It never creates one.

One brain, separate capabilities. Each bot is spawned with its own allowlist, chosen from functions the harness already has. Norma’s list is the full set, including creating bots. A savings bot’s list is only what it was given, such as asking Norma for a transfer. Memory cannot add a capability. All of a person’s bots read and write the same past.dev memory, so a preference saved by one is visible to the others. A memory is a fact or a preference, not a permission. Balances are looked up fresh, and a bank login stays in a vault the bot can name and cannot read.

Each bot has one conversation with the person, and it keeps going. There isn’t a second chat. Clearing the messages wipes that thread. The bot, its capabilities, and the shared memory stay.

The tables. Each person uses the same tables. A new bot adds rows. It does not get its own database.

  • bots. One row per bot, including that person’s Norma. The name, the owner, whether it is the main bot, who created it, and which version is active.
  • bot_versions. One row each time a bot is approved, kept forever. The model, the job in the words the person approved, and the approval rule in a form the code can enforce. A run uses one version. A later edit writes a new row.
  • bot_capabilities. The allowlist for that version, one row per function. The name has to be one the harness already has. Memory cannot add a row here.
  • messages. The one conversation, stored on the bot. This is the context the agent sees.
  • routines. A schedule on a bot, such as every payday, pointing at the version it should run. A routine can start the work. It still cannot move money on its own.
  • approvals. One row per move a bot wants Norma to make. The amount, the source, the destination, and whether the person approved, delayed, or changed it. Code moves the money only when this row says approved.

Memories are not in these tables. They go to past.dev under that person, with the writing bot’s id on each fact. Bank logins sit in a vault on the person. A version may be allowed to use a named login. It never receives the secret.