Export#
Through the API
Exporting and importing the agent definition is done through the API v1
(/api/v1/agentes); the screen doesn't have a button for it yet. Duplicating an agent
within the same workspace is done on the screen (Agents list › Duplicate).
An agent becomes a file with its entire definition: goal, instructions, tables, tools, knowledge bases, and team.
References are exported by name, never by internal identifier. An identifier from another environment means nothing at the destination — importing that way would create an agent that looks configured and doesn't work.
No credentials go into the file.
Import#
When importing, the platform looks up each reference by name at the destination and tells you, item by item, what it didn't find:
Table
gold_vendas_2024doesn't exist in this workspace · knowledge base "Políticas" not found
The imported agent is always created as a draft, with the pending items visible. An agent imported "successfully" that answers wrong because half of its references didn't resolve is the worst possible outcome.
And when there are two identical names at the destination, nothing is linked: picking one of the two would be invisible, because the name on screen would look right.
API#
There is an API for reading: listing agents, exporting the definition, querying runs. The technical description (OpenAPI) is available in the API itself.
Running an agent with an API key doesn't exist yet. The reason is the same permission rule as the other screens: every run happens with the permission of a person, and a key is not a person. Making the key pass as someone would work — and the audit log would start saying that person asked, when it wasn't them. We'd rather not have the feature than have a feature that lies.