How We Built Agent Mode for Genie

The lesson matters for anyone building agents over enterprise data, especially teams designing conversational analytics experiences with Genie. Simplicity, shared context, and production feedback loops drive better answers at lower cost. Teams evaluating how AI is changing the analyst's workflow will find broader context in why AI is rewriting the data analyst's job description.

The core idea: start with one agent, make it state assumptions, add lightweight planning only where needed, and turn runtime errors into targeted guidance the model can use to self-correct.

Why a single agent beats a supervisor of sub-agents

The initial multi-agent architecture looked tidy but failed in practice. Context went missing between hops, so discoveries like a customer's canonical name rarely carried forward. Each handoff added tokens, latency, and opportunities to hallucinate, while quality and cost suffered.

A single agent equipped with a small set of composable tools, primarily execute SQL and produce a response, kept context intact across steps. The model could plan end to end, including how to shape final visualizations, and it aligned with what foundation models already do well: reasoning inside a tool-calling loop. The practical rule: when steps have complex relationships, keep them within one agent rather than fragmenting work across sub-agents.

Getting consistency with stated definitions and a flexible checklist

Longer trajectories can drift, silently redefining key terms midway through an analysis. Forcing the agent to state assumptions and definitions up front, in plain language the user can verify, increases trust and provides a concrete specification the model can code against. It reduces inconsistency across multi-step queries and long conversations.

Open-ended questions like why revenue spiked benefit from early planning without reviving rigid orchestration. A lightweight checklist helps: the agent drafts initial hypotheses and tasks while context is clean, then executes SQL and updates the plan. It completes or cancels tasks as data validates or disproves assumptions, and it adds new lines of inquiry as patterns appear. Planning guides the work but never locks the agent into a brittle path.

Turning failures into targeted, in-run guidance

Production agents will hit timeouts, write bad joins, or misjudge data size. Generic failure messages force guessing. Specific, prescriptive feedback, such as query canceled due to timeout, consider selecting fewer rows, lets the agent self-repair in run. Even successful tool calls can steer behavior by explicitly noting truncation or result size, nudging the model toward outputs that fit a final report.

A progressive disclosure approach keeps the system prompt lean. Only reveal expert guidance when an error or edge case occurs. The loop becomes self-healing, catching and correcting issues before users see them, which improves reliability while controlling token use and latency.

Shipping smarter by resisting over-architecture

Good agent design rewards restraint. Start with one agent and a few deterministic tools. Make the model declare definitions so users can align on intent and the agent can code to a clear spec. Add lightweight, revisable planning where drift appears. Treat every production error as a new moment to teach the agent how to do better next time.