Configure an agent
The Configuration tab — instructions, model, knowledge bases, tools, approvals, credit ceiling and memory. Saved with Publish new version.
Nothing on this tab takes effect until you click Publish new version at the bottom; the agent
gets a new version number. Needs agent.update to edit and agent.publish to publish.
Brief and model
| Field | Notes |
|---|---|
| Instructions | The job description. Added after Ragenta's own base rules (an agent cannot be told to ignore the safety limits). Write it as you would brief a new hire: goal, when to search, when to stop, tone. |
| Model | Workspace default or a chat model your plan allows. |
| Temperature | Randomness. Low for work that must be consistent (extraction, classification), higher for drafting. |
| Max output tokens | Cap on each generation. |
Retrieval
- Knowledge bases — tick the ones the agent may search. All must share the same embedding model (the form blocks otherwise).
- Passages, Search mode (Hybrid / Vector only / Keyword only), Answer only from documents — same meaning as in Chat.
Tools
A checkbox list, 8 tools maximum. Each row has a description; two labels matter:
- changes things — the tool writes or sends to the outside (Call a connected system, Send an email, Save a document, Send from Gmail, Append to a Google Sheet, Post to Slack, Open a GitHub issue).
- Needs the X connection — requires a connection under Connections (Google, Slack, GitHub, Notion, your own API, a database) or a deployment-level integration.
MCP tools from the workspace's MCP servers appear here too, named server:tool.
The full list: Knowledge search · Web search · Fetch a URL · Call a connected system · Send an email · Save a document · Read a document image · Look at an image · Transcribe a recording · Speak text aloud · Read a spreadsheet · Create a spreadsheet · Open a page in a browser · Remember something · Search what you remember · Search Gmail · Send from Gmail · Search Google Drive · Read Google Calendar · Read a Google Sheet · Append to a Google Sheet · Post to Slack · Read a Slack channel · Search GitHub issues · Open a GitHub issue · Search Notion · Look something up in your database.
Safety and cost
| Field | Notes |
|---|---|
| Ask before it changes anything | On: before every changes things tool the agent pauses at awaiting_input and waits for your Approve / Decline in the Run tab. Turn it on for any agent that sends email or writes data. |
| Maximum rounds | 1–10. A round is one model turn plus the tools it calls. The cap that stops an agent looping forever. |
| Credit ceiling per run | The run stops when it reaches this. Set it low while experimenting. |
Memory
- Remember between runs — enables Remember something / Search what you remember.
- Whose memories — Shared by everyone (one memory for the agent) or Private to each person (a separate store per user; right for support and sales).
- Memories recalled per run — the 1–20 most relevant memories loaded into each run.
Memories are read back only by this agent and are always treated as data, never as instructions.