Your AI Agents Are Drowning in Bad Context

The essay recaps episode 52 with Michel Tricot, co-founder of Airbyte, the open source data integration platform. Its thesis, and the one behind Airbyte’s agent engine, is that agents fail because of data, and that many production agents are slow, expensive and unreliable for reasons unrelated to the model.

The opening example is a demo Tricot ran on the show: one agent task, finding all sales calls for a specific rep, done two ways. With a context store, the agent retrieved 45 calls in about a minute. Against the raw Gong API, which cannot filter by user, it paged through every call and took three minutes and 30,000 extra tokens. The essay notes that this was a single query and that the cost multiplies across a real workload.

Key points

• Teams that start with chunk, embed and retrieve end up annotating chunks with metadata until they have recreated a centralized data store the long way. The essay calls this the RAG death spiral.
• Tricot’s alternative drops embeddings. The context store leans on each system’s schema and on fuzzy matching to tell when records in Salesforce, Zendesk and Gong refer to the same thing.
• The essay says RAG is not dead, and that the fix for its default pattern looks like data engineering.
• Someone has to design how information reaches agents: its freshness, scope and governance. Tricot expects this “context engineer” work to be automated eventually, as prompt engineering was.
• Tricot’s framing: “Memory is a form of context. Context is just an understanding of the world.”

Disclosure: Episode 52 was sponsored by Galileo, the Season 3 presenting sponsor (read at 1:30).