Agents
An agent is a registered specification attached to a project. Its base type, prompt, enabled state, supported configuration, and allowed tools determine how it participates in the query workflow. The project-agent endpoints are listed in the Agents API reference.
Built-in Knowledge agent types
The Knowledge product registers knowledge_base, advisor, and
documentation agent types. They share retrieval and answer-generation
infrastructure, with different response behavior. Documentation version
filtering is currently a pass-through; selecting that type does not by itself
filter content to a particular product version.
The current retrieval pipeline uses vector search and BM25 full-text search, then merges and may rerank candidates. Typo correction is applied before keyword retrieval. The former separate fuzzy-search leg is no longer used.
Multiple active agents
When a project has multiple active agents, the workflow can use a supervisor to select an agent based on its routing description. With one active agent, the workflow dispatches directly to that agent. Routing depends on the registered agents and runtime configuration; it is not a guarantee that every project uses a multi-agent supervisor.
Configuration and tools
Agent configuration is layered from registered defaults and project overrides. The configuration drilldown endpoint reports the merge precedence. The API accepts prompt and configuration updates, but the shared query workflow does not apply every stored model parameter as a runtime override. Treat a setting as enforced only when its runtime consumer is implemented.
Available tools depend on the agent specification, enabled modules, and deployment settings. Do not assume that external web search or write-back actions are available to every agent. Review the project's allowed-tool configuration and test effects in the target deployment before enabling them.