PATTERN Cited by 1 source
Query language as agent tool¶
Query language as agent tool exposes a declarative query
language (typically SQL) as the primary tool surface for data
retrieval, instead of wrapping each REST endpoint as a tool.
Agents are strong at writing SQL, and SQL gives them
fine-grained control over what enters the
context window —
SELECT only needed fields, LIMIT to a few rows, COUNT /
GROUP BY server-side instead of retrieving raw samples
(Source: sources/2026-03-04-datadog-mcp-server-agent-tools).
Failure mode this replaces¶
V1 tool design: one tool per API endpoint
(get_logs_matching_filter, etc.). A user asks "which services
are logging the most errors in the last hour?". The V1 agent:
- has no aggregation tool → pulls some logs matching an
errorfilter, guesses a trend from the sample; often wrong. - worse, some agents brute-force it — repeatedly retrieving data until the context window fills.
With a SQL tool, the same question becomes one tool call of a
query like SELECT service, COUNT(*) AS c ... GROUP BY service
ORDER BY c DESC, returning small structured output.
Reported outcomes¶
Datadog reports on their eval scenarios:
- ~40% cheaper runs in some scenarios (fewer tokens to reach the answer).
- Higher correctness on trend / aggregation questions (exact counts, not sample inferences).
Cost of the pattern¶
- Significant implementation lift: "at the scale we operate at, traditional relational databases don't work". Requires a distributed query engine over the domain's storage (implicitly systems/husky-class, for Datadog).
- Surfaces more capability than necessary for the simple cases; pair with patterns/tool-surface-minimization so the surface area stays manageable.
- Errors must be specific to be useful (patterns/actionable-error-messages); "invalid query" traps agents in retry loops.
Applicability¶
- Strong fit: data retrieval where the agent asks aggregation, filtering, or ranking questions; domain with stable schema and affordable query-engine coverage.
- Weaker fit: action tools (create / modify / delete) — SQL collapses retrieval, but side-effecting operations remain separate named tools.
Seen in¶
- sources/2026-03-04-datadog-mcp-server-agent-tools — adopted as the primary observability-data retrieval surface on Datadog's MCP server after V1 sample-and-infer patterns failed.