AI Assistant — Your Platform Co-Pilot

The AI Assistant is a context-aware chat panel embedded in every page of the app. Ask it anything about your data, have it create or update entities, or delegate platform operations — all in natural language.

Scenario: You're reviewing last week's experiments and ask the assistant: "How is the weekly digest project doing? How much did it spend this month?" The assistant queries your live data and responds with the current run count, success rate, and exact credit usage.

Where to find it

The assistant panel is accessible from every page via the floating button in the bottom-right corner. Click it to open a full-width chat interface. It opens in the context of whatever page you're currently viewing.

Context awareness

When you open the assistant on a specific page, it automatically receives the entity's context:

On an experiment page: knows the experiment ID, status, and stage history
On a project page: knows the project's schedule, recent runs, and budget
On an agent page: knows the agent's config, skills, and health status
On the dashboard: receives the KPI summary (runs, costs, agent health)

You can reference the current entity naturally: "Retry this experiment from step 3" — the assistant knows which experiment "this" refers to.

What the assistant can do

The assistant has access to 28 internal tools, role-gated by your team role:

Category	Role required	Examples
Read	Any	List experiments, get agent status, query budget, search audit log
Create / Update	Member+	Create agents, start experiments, update skills, trigger project runs
Control	Member+	Pause, resume, retry, kill experiments
Delete / Destructive	Admin+	Delete agents, remove team members, purge semantic cache

Under the hood: the tool loop

When you send a message, the assistant sends it to the LLM along with a list of available tools. The model decides which tools to call, calls them (against your live data), and loops until it has enough information to answer. Complex queries may invoke 4–6 tools internally.

The assistant uses a sliding context window of 30 messages (~50k tokens). Older messages are summarised automatically as the conversation grows, so you never lose context.

Example prompts

"Show me all failed experiments from the last 7 days"

"Create a new agent called "Lead Qualifier" with a sales-focused goal"

"How much credit did we spend on the blog content project this month?"

"Retry the last failed experiment from the scoring step"

"What agents are currently unhealthy?"

"List all pending approval requests and their deadlines"