Agent Stack¶
sqldbagent uses LangChain v1’s create_agent(...) interface over LangGraph runtime primitives, with sqldbagent-owned tools, prompts, middleware, and persistence boundaries.
Design Rules¶
Use sqldbagent’s service layer, not raw SQL toolkit calls, as the primary execution path.
Seed agent state from stored snapshots so agents start with durable schema context.
Keep safe SQL and retrieval as explicit tools.
Use Postgres checkpointing for durable threads and in-session memory fallback for lightweight dashboard runs.
Let LangSmith observe the same surfaces instead of adding a second tracing model.
Middleware¶
The default middleware stack currently covers:
state seeding from stored snapshot context
dynamic prompts
tool error shaping
tool digest compression
optional todo middleware
optional HITL middleware
optional summarization middleware
model and tool call limits
Surfaces¶
langgraph devvialanggraph.jsonStreamlit dashboard via
sqldbagent dashboard serveFastMCP via
sqldbagent mcp servedirect Python runtime usage through
sqldbagent.adapters.langgraph
LangSmith¶
Dashboard turns are wrapped in a LangSmith tracing context when tracing is enabled. LangGraph runtime usage can inherit the same .env-driven LangSmith configuration because langgraph.json points at the repo .env.