Skip to main content
Everything about an agent lives on one page. Your Agents puts a live test chat on the left and the agent’s configuration on the right, so you can change something and immediately ask a question about it.

​
Creating one

Click Create Agent. You get a draft straight away — the Draft chip stays on it until you publish.
One draft at a time. If you already have an unfinished draft, creating a new agent replaces it, and everything in it goes — including any sources you added. AgntLabs names what it is about to delete before doing it.
Drafts do not count against your plan’s agent limit. Only published agents do.

​
The configuration column

Four fields, one Save for the whole block:Role and Goal shape how the agent behaves, not just how it is labelled.
Free-text rules for how your agent should behave — what to always mention, what to never promise, how to handle a specific kind of request.The row shows a short preview so you can see whether instructions exist without opening the panel.
What the agent has read. The closed row summarises it — total sources, plus how many are still unpublished or failed to read.See Knowledge base.
Storage used, source counts and other stats for this agent.
Below these you will see Tools, Workflows, Triggers, Schedule and Guardrails marked Soon. They are placeholders for work in progress — they do nothing yet.

​
The knowledge base summary

The closed Knowledge base row is deliberately precise. “18 sources” on its own would suggest eighteen sources your agent can answer from, which is not true while some are unpublished or unreadable. So it reports them separately:
  • Total — every source in the list
  • Unpublished — added, but with no vectors behind it yet. Press Publish.
  • Failed — could not be read at all. A website that would not crawl still occupies a row.
A failed source counts in the total. If the numbers look higher than the answers your agent can give, open the panel and check for failures.

​
Testing

The left pane is a real conversation with the selected agent, using the same pipeline your customers hit — the same retrieval, the same actions, the same deflection behaviour. Worth trying:
  • Your top questions. Are the answers right?
  • Off-topic questions. It should decline, not improvise.
  • Vague questions. Does it ask a sensible follow-up?
  • Each enabled action. Ask naturally — “can I book a call?” — rather than reciting keywords.
Your own test chats are counted separately from customer traffic in Analytics, so testing does not distort your quality metrics.

​
Publishing

Publish lives in the Knowledge base panel. It:
  1. Reads every unpublished source
  2. Splits the content into chunks
  3. Creates embeddings and stores them for this agent
  4. Applies any staged edits or removals to live sources
  5. Turns a draft into a live agent
Until you publish, a new source is inert.
Save your Identity fields before publishing. Publishing refreshes the page, which discards unsaved text in those inputs. AgntLabs blocks Publish while they are dirty.
If some sources fail, the rest still publish. The failed ones stay in the queue so you can retry.

​
Agent limits

Drafts are free. See Plans & Billing.

​
Shared agents

If you have been added to someone else’s account, their agents appear in your list alongside your own. A Viewer sees a “View only” banner at the top of the configuration column and gets no editing controls at all — no Save, no Upload, no Add source. That is intentional: a button that would fail is worse than no button. See Team members.