Back to Changelog
v0.11.0
March 17, 2026

from_db(), Watch System & Data Plugins

AddedChangedFixedRemoved

#[0.11.0] - 2026-03-17

This release introduces three major capabilities: Agent.from_db() for zero-config database agents, the @agent.watch() system for real-time data monitoring, and a new suite of data plugins (SQLite, Transformer, Data Quality). The agent core has been significantly restructured for clarity, and the graph backend now supports custom distributed implementations.

#Added

  • Agent.from_db() — Database Agent Factory

    • Create a fully configured agent from a connection string or plugin instance with a single await call
    • Automatically connects, discovers the schema, generates a context-aware system prompt, and registers the appropriate query tools
    • Infers the data domain from table names to produce a more natural default agent name and prompt context
    • Numeric column sampling provides the LLM with value distributions; PII columns are redacted by default (redact_pii_columns=True)
    • Schema discovery results are cacheable with a configurable TTL (cache_ttl=) and drift-detected on subsequent calls
    • Opt-in lineage=, memory=, and history= parameters auto-create and attach the corresponding plugins
    • read_only=True (default) restricts the agent to query tools only
    • toolkit="analyst" (default) registers six analytical tools: pivot table, correlation, anomaly detection, entity comparison, similar-record search, and trend forecasting
    • Supports PostgreSQL, MySQL, SQLite, and MongoDB connection strings

  • Analytical Toolkit (daita.agents.db.tools)

    • Six plug-and-play tools available via Agent.from_db() or manual registration:
      • pivot_table — group and aggregate tabular results in-memory
      • correlate — compute Pearson correlations between numeric columns
      • detect_anomalies — flag statistical outliers using IQR
      • compare_entities — diff two entities row-by-row across a key column
      • find_similar — rank records by similarity to a reference entity
      • forecast_trend — linear trend projection over a time-series column
    • All tools work on any tabular data returned by a previous query tool

  • @agent.watch() — Real-Time Data Monitoring

    • New decorator for defining data-driven event handlers on a running agent
    • Supports polling sources (database plugins, async callables) on configurable intervals ("30s", "5m", "1h", "2d")
    • Condition evaluation via SQL WHERE strings or Python predicates; threshold callbacks for numeric comparisons
    • on_resolve=True fires a second event when a triggered condition clears
    • Streaming source protocol (WatchSource) for push-based integrations
    • WatchEvent carries the current value, previous value, timestamp, source type, and resolution flag
    • Exported from top-level package: from daita import WatchEvent

  • SQLite Plugin

    • New SqlitePlugin with full async support via aiosqlite
    • Tools: sqlite_query, sqlite_list_tables, sqlite_get_schema, sqlite_inspect, sqlite_execute
    • Supports in-memory databases and file-based databases
    • Compatible with Focus DSL pushdown and Agent.from_db()
    • Install with: pip install 'daita-agents[sqlite]'

  • Transformer Plugin

    • New TransformerPlugin for declarative in-memory data transformations
    • Tools: filter rows, rename columns, cast types, aggregate, pivot, join, deduplicate, and compute derived columns
    • Operates on data returned by other tools without additional database access
    • Install with: pip install 'daita-agents[data]'

  • Data Quality Plugin

    • New DataQualityPlugin for validating datasets against configurable rule sets
    • Built-in rules: null checks, uniqueness, range bounds, pattern matching, and referential integrity
    • Returns a structured report with per-column pass/fail counts and failing row samples
    • Install with: pip install 'daita-agents[data]'

  • Custom Graph Backends

    • register_backend_factory() allows replacing the default LocalGraphBackend with any implementation that satisfies the GraphBackend protocol
    • Enables distributed or cloud-backed graph storage for Lineage and Catalog plugins in multi-agent deployments
    • GraphBackend protocol is @runtime_checkable; the local NetworkX/JSON backend remains the zero-config default

  • Connection Retry & Observability

    • BaseAgent gains a configurable connection retry loop with exponential backoff
    • Per-retry events are emitted to the trace system, surfacing retry counts and failure reasons in the dashboard

#Changed

  • Agent Core Restructured

    • agent.py refactored: initialization, tool dispatch, plugin management, and execution are now in clearly separated internal layers
    • BaseAgent slimmed down — lifecycle hooks and retry infrastructure consolidated; watch startup is deferred until after _running=True
    • Plugin connect() lifecycle formalized: plugins now expose a richer interface including health-check and disconnect hooks
    • AgentConfig extended with retry policy fields surfaced in the constructor

  • Focus System: SQL Backend Improvements

    • SQL pushdown backend rewritten to handle a broader set of DSL constructs natively
    • GROUP BY with aggregates now pushes correctly for all supported dialects
    • Dot-notation field access and unsupported expressions fall back gracefully to the Python evaluator
    • evaluate_remaining now skips projection when aggregate columns are absent from result rows, preventing empty-result regressions

  • Catalog Plugin: Distributed Consistency

    • Internal graph writes use the registered backend factory, so multi-node catalog deployments share a single graph store
    • Schema comparison and diagram export paths updated to work through the GraphBackend protocol

#Fixed

  • **Agent **kwargs Forwarding**

    • Extra keyword arguments passed to Agent() are now forwarded correctly to the underlying LLM provider, fixing silent drops of provider-specific options

#Removed

  • run_detailed() — Replaced by run(detailed=True)

    • agent.run_detailed(prompt) has been removed
    • Use agent.run(prompt, detailed=True) to get the full execution dict (result, tool_calls, iterations, tokens, cost, elapsed time)
    • Plain agent.run(prompt) continues to return the result string as before

  • RowAssertion

    • Removed RowAssertion from daita.core.assertions and from the top-level package exports
    • Use column-level assertions (NotNullAssertion, UniqueAssertion, RangeAssertion, etc.) or DataQualityPlugin for row-level validation