Skip to content
CompanyWizard

The agent model

All

An agent carries out a recurring task on its own. So that this happens predictably and safely, every agent is assembled from five building blocks. Anyone who knows these understands any agent — and can test it safely.

The five building blocks source, reasoning, tools, trigger, identity feed into the agent run, which produces a dry run or a real run. The five building blocks source, reasoning, tools, trigger, identity feed into the agent run, which produces a dry run or a real run.

The five building blocks

Section titled “The five building blocks”
  1. Source — where the agent takes its input from: a vault (knowledge area), an email mailbox, a working folder (desktop only) or none (just a prompt). A “scope” narrows down which part of the source applies.
  2. Reasoning — the instruction (system prompt) plus optional Skills that shape the approach. An optional result schema turns the final answer into a JSON result.
  3. Tools — the tools the agent is allowed to use. Each is marked as read (READ) or write (WRITE) — decisive for the risk assessment. A storage target (vault folder or working folder) sets where writing tools may write.
  4. Trigger — when it runs: on a schedule (cron), manually, as a dry run or triggered by a chat message.
  5. Identity — under which user the agent runs. This determines its rights (RBAC): an agent sees and does only what that user would be allowed to.

How a run proceeds

Section titled “How a run proceeds”

The agent run gathers input from the source, reasons with the model, instruction and Skills, and uses the permitted tools along the way. In a dry run, all write tools are blocked — the agent only shows what it would change (“Planned changes”). Process items one by one is suppressed as well: it only reads, but it triggers paid model calls. → Tools: Process many files or mails one by one Only the real run executes write actions. Every run is logged.

Safety principle: Reading is harmless, writing needs trust. Combine WRITE tools with automatic triggers only after a successful dry run.

Section titled “Related”