Chain of Thought | AI Agents, Infrastructure & Engineering Context Poisoning Is Killing Your AI Agents: How to Stop It
Mar 25, 2026
Michel Tricot, CEO and co-founder of Airbyte, a data integration leader building an agent engine and context store. He argues context poisoning—not models—is why agents fail. Live demos compare raw API calls to a context store, showing massive token and time savings. Discussion covers why RAG alone falls short, entity tracking across SaaS systems without embeddings, and the new role of context engineering.
AI Snips
Chapters
Transcript
Episode notes
Context Enables Agents To Reason Across Systems
- Context, not raw data volume, is what enables agents to form relationships across systems and act intelligently.
- Michel Tricot explains agents can map equivalent fields (e.g., position vs job title) when given descriptive metadata rather than dumping all warehouse rows.
APIs Reintroduce Data Access Complexity
- Hitting APIs directly recreates access complexity warehouses solved, causing rate limits, missing filters, and brittle agents.
- Michel describes agents failing because APIs often don't support the efficient queries agents need, forcing costly workarounds.
Preprocess And Enrich Data Before Feeding Agents
- Build post-processing and metadata enrichment before feeding LLMs to avoid context pollution and wrong results.
- Michel advises attaching structured metadata to transcripts and records so agents return relevant snippets instead of bloated lists.