The Assistant

The assistant

Auton8 ships an in-app agent — the chatbot in the right-hand sidebar. It is not a general chatbot bolted on: it has tools that call the platform's own APIs, so it can answer questions about your data.

What it is for

It introduces itself as a diagnostic assistant, and that is an accurate description of its centre of gravity: helping administrators and operators understand the state of records, workflows, notes, dashboards, permissions and system issues by reading data through tools and citing what it used.

What it can do

The tool surface is broad, and grouped roughly as:

• Look things up — records, record types, notes, dashboards, datasets, data stores, workflow executions, permissions, the user directory, notifications
• Explain — a workflow model, an AQL query, why a permission resolves the way it does, a system issue
• Manage — records and record types, record edges, notes, dashboards, datasets, data stores, saved queries, permissions, plugins, site settings
• Operate — workflow executions
• Reach out — fetch a URL, or run a web search, when configured
• Notify — send notifications

Plugins can contribute their own skills, so an installed plugin can extend what the assistant is able to do.

What it cannot do

Two limits are structural, not settings:

Every tool runs as you. The agent has no privileges of its own. A tool call is authorized exactly as the same action would be from the UI, so the assistant cannot read or change anything you could not. If it says it cannot see something, the answer is a grant, not a prompt.

Tool results are data, not instructions. Text inside a record value, a BPMN annotation, or a system-issue summary is untrusted input. The agent is instructed never to follow instructions embedded there — a defence against someone planting an instruction in a field for the agent to find.

It also has no write access to workflows or system issues; asked to change one, it will explain that and point you at the UI.

Configuration

Two admin surfaces, deliberately separate:

Site Configuration → Chatbot → Chatbot Settings controls capabilities — what the agent is allowed to do. Changes apply on the next message.

Site Configuration → Chatbot → Models is the model catalogue. Models carry their context windows, which the agent loop reads, so the catalogue is not cosmetic: a wrong entry produces wrong truncation behaviour.

The provider itself is an external connection (Site Configuration → Sitewide → External Connections), holding the endpoint and API key. Keys are encrypted at rest, and outbound requests are constrained to an allow-list of hosts per connection kind.

There is no bundled model and no default endpoint — the assistant does nothing until you configure a provider.

Conversations

Conversations are listed, renamed, compacted and deleted from the sidebar. The full transcript is stored; the audit log links to it by id rather than copying prompt or response text, so the event stream carries ids, lengths, token counts and tool names, not content.

Page context

The assistant can be aware of the page you are on, so "what is this?" has a referent. That context is provided by the page itself, which is why the answer quality varies by surface — a page that publishes rich context gets better answers than one that publishes none.