The 30-Day Cliff
In early 2025, we noticed a pattern across multiple client deployments: initial performance metrics looked excellent at go-live. Two weeks later, still good. At the 30-day mark, something started to slip. Not dramatically — maybe 8–12% accuracy degradation. But in regulated workflows, 8–12% matters.
We instrumented 14 production deployments to understand the degradation pattern. The results were consistent enough to call it a law: without active context engineering, most AI agents will degrade measurably by day 30 and significantly by day 60.
This article explains what causes it, how to detect it, and the specific patterns that maintained 92% performance retention at day 90.
---
What Causes Post-Launch Degradation
The causes fall into four categories, each with distinct signatures:
1. Context Window Pollution
Agents that accumulate conversation history or working memory without active management will eventually pass context that confuses rather than helps the model. We call this "context pollution."
In a healthcare prior auth agent we deployed, the agent's working memory grew from an average of 1,200 tokens at go-live to 8,400 tokens by day 45. The additional context included: - Resolved workflow states that were no longer relevant - Error recovery artifacts from earlier failed attempts - Redundant retrieved chunks from the same source documents
At 8,400 tokens of working memory, the model's attention was diluted across a larger context than it was optimized for. Accuracy on the core task dropped from 97.1% to 89.3%.
Fix: Implement a context compression agent that runs before every main agent invocation. The compression agent's only job is to summarize and prune the working memory to a target token budget we use 2,000 tokens for most workflows. The summary preserves the semantic content while discarding resolved states and redundant information.
2. Retrieval Drift
RAG systems retrieve context based on embedding similarity. Over time, the knowledge base grows new documents added, but the retrieval parameters top-k, similarity threshold, chunk size stay fixed. What was an appropriate retrieval config at go-live may return increasingly noisy results as the corpus grows.
We measured this explicitly: in a legal contract review system, retrieval precision fraction of retrieved chunks that were actually relevant dropped from 87% at go-live to 64% by day 60 as the case law corpus grew from 12,000 to 31,000 documents.
Fix: Monitor retrieval precision separately from end-to-end accuracy. We run a weekly retrieval quality evaluation using a held-out evaluation set of query-document pairs with known relevance labels. When precision drops below 75%, we trigger a retrieval parameter tuning sprint.
3. Distribution Shift
The documents and inputs your agent sees in production drift from the documents it saw during training and validation. This is especially pronounced in: - Seasonal workflows healthcare prior auth requests sh