Sponsored by: Pogo | You're Overthinking It: What Analyzing $500B in Consumer Spend Taught Us About Building AI Agents

Dom Wong, Luke Morgan-Scott16 Jun 2026

Problem

Pogo built an AI research agent that must answer any brand question across billions of purchase records and millions of interview transcripts. The first version was a multi-agent system: a supervisor over a SQL agent, a Python agent, an analyst and a report generator. Errors cascaded between agents, each saw only its own work, and fixing one broke another. The team spent its time on handoffs instead of the user's question.

Example from the article: a McDonald's team asks who buys the new Big Arch Burger, whether buyers like it, and whether the CEO's viral TikTok helps or hurts. Answering needs both purchase queries and interview analysis.

Solution in brief

Pogo collapsed the system into one agent that runs code, reads the output and corrects itself, keeping full context. Every analysis question is treated as a coding task, and domain expertise is loaded on demand through skills rather than extra agents.

Questions this case answers
  • How to build an LLM agent that answers analytical questions over large transaction data and unstructured interview transcripts in one loop
  • How to decide between a multi-agent system and a single agent for an analytics assistant that must keep full context
  • How to build an LLM agent that answers analytical questions over transaction data and interview transcripts
  • How to decide between a multi-agent system and a single agent for data analysis

Tools and methods

  • Single-agent loop (run code, see output, self-correct) — replaces the supervisor and sub-agents
  • Code execution environment — safe sandbox for SQL and Python with saved files and logs
  • Skills — title and short description in context; full instructions loaded only when needed
  • Automated validation scripts — gate the output before the user sees it
  • Purchase transactions and interview transcripts — the quantitative and qualitative layers
Continue reading — it's free

Google or a one-time email code. No password.