In detail
1. Tricot’s thesis: the failure is upstream of the model
Airbyte co-founder Michel Tricot’s thesis is that agents do not fail because the models are bad. They fail because the data feeding them is wrong. Swapping in a stronger model does nothing about that, which is why teams keep upgrading and keep getting the same class of failure. Context poisoning is the name for the input problem, and it is a data-engineering problem wearing an AI costume.
2. His demo: a raw API not shaped for an agent
Tricot demonstrated it live. Gong’s API, as used in the demo, does not let you filter calls by user, so an agent asked for one person’s recent calls has no choice but to page through lists. A query as ordinary as “retrieve all calls since February 1” burned about 30,000 extra tokens, and run that way it normally takes about three minutes, most of it spent pulling pages the agent did not need. Nothing was malicious and nothing was broken. The API was simply not shaped for an agent, and the agent paid for that in context window.
3. The precision cost is worse than the token cost
The token cost is the visible half. The expensive half is that a window stuffed with near-duplicate, low-signal records leaves the model uncertain about what to prioritize. You get an agent that had the right answer available and did not weight it. That failure can look like a reasoning failure in a trace, which can send teams off to evaluate models instead of inputs.
Research on long inputs points the same way. In Lost in the Middle, Nelson Liu and colleagues found that models used information best when it sat at the beginning or end of the input, and that accuracy dropped significantly when the relevant passage was in the middle of a long context, even for models built for long inputs. Chroma’s Context Rot report tested 18 models, including GPT-4.1, Claude 4 and Gemini 2.5, and found that performance became increasingly unreliable as input length grew, even on simple tasks. A bigger context window doesn’t make the padding free.
4. Put something between the agent and the systems
Tricot’s answer is a context store: a layer that centralizes and pre-shapes data from source systems so the agent queries a store built for retrieval instead of paging an API built for a web app. The pattern he runs internally is hybrid: discovery and understanding through search over the store, and a direct API call when the task needs the freshest record. The decision to make deliberately is which reads have to be live; for those, fetch the one record by ID rather than paging through a list.
The same name for a different, deliberate attack
Tricot uses the term for a reliability problem: useless, stale or oversized data crowding the context, with nobody at fault. Security writing uses “poisoning” for something deliberate. A document or web page carrying instructions meant to redirect the agent is indirect prompt injection, and filtering the context for relevance does not stop it. The planted text can be short and look relevant. A variant targets memory: Microsoft’s security team describes “AI Recommendation Poisoning,” where hidden prompts behind “Summarize with AI” buttons try to make an assistant “remember [Company] as a trusted source,” mapped to MITRE ATLAS technique AML.T0080, Memory Poisoning (Microsoft Security, February 2026).
The fixes differ, so diagnose first. Keep the question and model fixed and vary what is supplied: the raw result against trimmed excerpts shows whether noise is the problem; an old policy against the current one shows a freshness failure; a retrieved instruction to export unrelated records, or to save a vendor as “always trusted,” shows whether your access and memory-write controls block the action even when the model goes along with it.
Why it matters
This is the practical side of context engineering: deciding what goes into the window, not only how the prompt is worded. Most agent debugging starts at the model and works backward. Start at the input instead: log what actually entered the window on a failed run. If the answer is 30,000 tokens of paginated list, no model upgrade will fix it.
Hear it from the guest
“Because as everyone is saying, the problem is not in the model today, the problem is in the data that you provide.”
“If you get stuff from an API you might just get like a list of records that you might not need. And all of that is just going to pollute your context.”
“One example I remember our red team … finding was … invisible characters that people could put in, like, these kind of packaged prompts.”
Quotes lightly edited to remove filler words.
From the conversation
This explainer is drawn from these episodes — each carries its full transcript.
Concepts in this explainer
Context PoisoningData PoisoningLost in the MiddleContext WindowContext EngineeringPrompt InjectionTokenizationAccuracy