Working with agents

An agent is a colleague, not a search box. Each one covers a discipline — Hardware, Quality, Sales, Finance and so on — and each has its own brief, its own way of working and its own documents to look after. You get better answers when you treat them like specialists and ask the right one.

Pick who answers

The coloured pill at the bottom of the chat, inside the composer, decides which agent replies. Click it and you get:

• ● Roundtable (all) at the top — every enabled agent at once, described in The roundtable;
• a separator;
• every agent available in the area you are currently in.

You can also click an agent in the sidebar. That starts a fresh conversation with them.

Ask the owner of the answer

Ask Supply Chain about lead times, Regulatory about which directives apply, Sales about what you have sold. Asking the wrong agent still gets you an answer, but a thinner one, because each agent's brief and tools are built around its own field.

Have a conversation

1. Select the product (Product Development only) and the agent.
2. Type your question and send it.
3. Watch the agent work. It shows you its steps as it goes — what it is reading, which tool it is calling and what came back.
4. Read the answer, then keep going in the same thread. The agent remembers everything earlier in the conversation.

Conversations are saved automatically and listed in the sidebar under the agent's name. + New starts a clean one — do that when you change subject, because a long thread about something else costs credits and muddies the answer.

Choosing a model

The second picker in the composer selects the language model, with its credit cost shown as a multiplier such as (1x) or (3x).

• Use the cheap model for lookups, summaries and quick questions.
• Use the expensive model for design reasoning, tricky trade-offs and anything you are going to publish.

Which models are on the list is set by your organisation under Settings → AI.

When an agent writes a document

Agents do not only answer — they write. When one changes a file, the canvas opens on the right and shows you what happened. There are two behaviours:

• Suggested changes appear as an inline diff you accept or reject.
• Direct edits are written straight into the document, and the canvas reloads to show the new text.

Either way nothing is published. The change sits as a draft until you publish it. See Documents and the canvas and Publishing and reviews.

Escalating a question

Sometimes you realise halfway through that a question is bigger than one discipline. Next to the agent pill there is an ↗ Roundtable button. It takes your last message across to the roundtable, tagged with where it came from, so all disciplines can weigh in without you retyping anything.

Looking under the hood

The ⋯ menu in the chat header opens two things:

• Execution plan — the phased plan for the product, showing which discipline is working on what and what has to be approved before the next phase unlocks.
• Agent settings — the selected agent's Mission Stack (the SOPs and instructions it follows) and its System Prompt. Worth reading once, so you understand why an agent answers the way it does.

Troubleshooting

Problem	Likely cause
The agent invents a number instead of looking it up	The system it should read is not connected. Check Settings → Integrations — see Integrations and your ERP.
The answer ignores your company context	The company profile is empty. Fill in Settings → Company.
The agent will not write to a document	Product Development agents need a product selected. Pick one in the sidebar first.
Answers stop coming	You may be out of credits. Check the balance in the top bar.

Next: The roundtable →