Prompt caching
Prompt caching for Claude and Gemini agents
An agent resends its instructions, its tools and the whole conversation on every turn. Claude, and Gemini through OpenRouter, only serve that repeated part from their cache when the request asks for it. Cruq AI now asks, with no setup on your side.On a six-turn Claude Sonnet run with a calculator tool, cost went from $0.063 to $0.024, and to $0.016 when the agent had run in the previous five minutes. Each model call in a run’s trace shows how much of its input came from the cache, for example3.1k in (97% cached).Keep anything that changes between runs, like today’s date or a customer’s
name, out of the agent’s instructions and pass it as run context, so the
instructions stay cacheable.Prompt cachingEvals and traces
Check the judge, and rubrics with dimensions
An LLM judge’s score is only as good as its reading of your rubric. On a comparison’s page, Check the judge lets you score some results yourself, 1 to 5. After five, it tells you how far the judge is from you, whether it leans generous or harsh, and which results you disagree on most. The judge’s score stays hidden until you give yours.llm_judge scorers can also list up to six dimensions, such as accuracy and
tone, each with written levels. The judge scores each one separately, so you
can see why an answer lost points.Checking the judge ·
Rubrics with dimensions ·
cruqai evals review <experiment-id>Eval cases suggested from real runs
Once an agent has traffic, open a suite and click Suggest from runs. Cruq AI picks real requests from the last 30 days: one for each group of similar ones, plus a run that failed, one where a guardrail tripped, one where the agent asked a person, one that used a rare tool and one unusually long run. Each comes with starting scorers for you to review before you add it.Cases from real runs ·cruqai evals suggest <suite>Every model call in the trace, and OpenTelemetry export
Runs now record each call the agent makes to its model: the model, tokens, cost, how long it took and the time to the first token. They appear between the tool calls on a run’s detail page.GET /v1/runs/{id}/otel returns any run as OpenTelemetry trace data, using
the GenAI semantic conventions, so you can load it into Langfuse, Phoenix,
Grafana Tempo or any OTLP collector.Model calls ·
Export to OpenTelemetry ·
cruqai trace <run-id> --models / --otelMCP
MCP server at mcp.cruq.ai, with sign-in
Addhttps://mcp.cruq.ai to Claude, ChatGPT, Cursor, Codex, VS Code or any
MCP client and sign in with your Cruq AI account. No API key to copy. From the
chat you can run your agents, read their runs and build new ones.