modelBitrequired
The LLM model id backing this agent
typeBitTypes
enum "Llm", "Vlm", "Tts", "Stt"...default "Other"
metaMap<string, Metadata>
default {}
*Metadatamap value
descriptionstringrequired
long_descriptionstring | null
release_notesstring | null
tagsArray<string>required
+11 more fields
authorsArray<string>
default []
repositorystring | null
default null
+14 more fields
model_display_namestring | null
Model display name
max_iterationsinteger:uint64required
Maximum number of iterations/tool calls before stopping
format uint64min 0
system_promptstring | null
System prompt for the agent
toolsArray<Tool>
Registered tools (function calling schemas for non-function tools)
default []
itemsToolarray item
typeToolTyperequired
enum "function"
functionHistoryFunctionrequired
parametersHistoryFunctionParametersrequired
function_refsMap<string, string>
Function references (node_id -> node_name mapping)
These are converted to tools at execution time to keep data slim
default {}
mcp_serversArray<McpServerConfig>
MCP servers with optional tool filtering
default []
itemsMcpServerConfigarray item
MCP server registration with optional tool filtering
uristringrequired
URI of the MCP server
tool_filterarray | null
Optional tool filter - if None, all tools are used
If Some, only tools in this set are used
thinking_enabledboolean
Whether the thinking tool is enabled
default false
historyanyOf (2)
Optional conversation history to initialize with
variant 1Historyvariant
messagesArray<HistoryMessage>required
itemsHistoryMessagearray item
stream_optionsanyOf (2)
variant 1StreamOptionsvariant
+14 more fields
datafusion_contextsArray<DataFusionContext>
DataFusion sessions for SQL-based data analysis
Multiple sessions can be added to give the agent access to different data sources
default []
itemsDataFusionContextarray item
DataFusion session context for SQL-based data analysis
session_cache_keystringrequired
Cache key to look up the session in ExecutionContext.cache
descriptionstring | null
User-provided description of what this data represents
e.g., "Sales data from 2020-2024 including customer demographics"
table_descriptionsMap<string, string>
Per-table descriptions for better LLM understanding
Key is the table name, value is description
default {}
example_queriesArray<string>
Example SQL queries that work well with this data
default []
table_schemasMap<string, string>
Auto-discovered table schemas (populated at runtime)
Key is table name, value is schema description
default {}
infinite_contextboolean
Enable infinite context mode with automatic context window management.
When enabled, applies the selected context management strategy.
default false
context_management_modeContextManagementMode
Strategy for managing context when it exceeds the token budget.
- Truncate: Sliding window, removes oldest messages (fast, no extra cost)
- Summarize: LLM compresses old messages (preserves info, adds latency/cost)
default "Truncate"
variant 1constvariant
Sliding window truncation - removes oldest messages to fit budget.
Fast, deterministic, no extra API costs. May lose important early context.
const "Truncate"
variant 2constvariant
LLM summarization - compresses old messages into a summary.
Preserves key information but adds latency and API cost.
const "Summarize"
max_context_tokensinteger | null
Maximum tokens to retain when truncating history in infinite context mode.
Defaults to 32000 tokens if not specified. Only used when infinite_context is true.
format uint32default nullmin 0
lazy_function_refsArray<LazyFunctionRef>
Lazy function references backed by a vector DB index.
At execution time the agent can search this index to dynamically discover
and load only the tools it actually needs, keeping the context window lean.
default []
itemsLazyFunctionRefarray item
Reference to a lazy function tool index stored in a vector DB.
Allows agents to do hybrid search over a large pool of tools at execution time
instead of loading all tool schemas into the context upfront.
db_cache_keystringrequired
Cache key used to look up the LanceDB connection
lazy_embedding_modelanyOf (2)
Embedding model shared across all lazy function tool indexes.
The model's cache key is encoded into the vector DB table name, so
swapping the model automatically uses a fresh table (old embeddings are abandoned).
variant 1CachedEmbeddingModelvariant
model_typeBitTypesrequired
enum "Llm", "Vlm", "Tts", "Stt"...
memoryanyOf (2)
Persistent memory configuration. When set, the agent gains built-in
`_memory_search`, `_memory_store`, and `_memory_compress` tools to
autonomously store, recall, and compress observations across conversations.
variant 1MemoryConfigvariant
databaseNodeDBConnectionrequired
embedding_modelCachedEmbeddingModelrequired
model_typeBitTypesrequired
enum "Llm", "Vlm", "Tts", "Stt"...
max_context_tokensinteger:uint32required
format uint32min 0
recall_strategyRecallStrategyrequired
enum "RecentFirst", "Relevance", "Hybrid"
recall_top_kinteger:uint32required
format uint32min 0
+2 more fields
