Overview
When context passes through multiple agents, important details can be lost or distorted. Each handoff increases the risk of losing key information, especially in long, multi-step conversations.
Drift Mechanisms
Summarization Loss
Each agent may summarize context before passing it on:
Original: "Client prefers blue, but will accept navy or sky blue,
NOT royal blue which they had issues with last time."
After 3 handoffs: "Client prefers blue."
Priority Inversion
Less important information displaces critical context due to token limits.
Semantic Shift
Meaning changes subtly through paraphrasing:
Original: "Handle with care" (fragile item)
Drift: "Handle with care" (sensitive topic)
Temporal Confusion
Agents lose track of what happened when:
"They said they'd call back" - Was this yesterday? An hour ago? In the original request?
Impact Scenarios
Customer Service
Customer explains complex issue to Agent A. By the time it reaches Agent D, the core problem is obscured by later conversation.
Research Pipeline
Key constraints from the original request are lost, leading to research that doesn't address the actual question.
Multi-Step Reasoning
Intermediate conclusions drop critical caveats, leading to overconfident final answers.
Detection Approaches
Context Checksum
Include critical facts in a structured summary that's validated at each step:
{
"core_requirements": ["blue preferred", "not royal blue"],
"constraints": ["budget under $1000"],
"checksum": "abc123"
}
Semantic Similarity Tracking
Compare context embeddings before and after each handoff:
if cosine_sim(original_context, current_context) < 0.85:
alert("Significant context drift detected")
Key Entity Tracking
Ensure all mentioned entities remain in downstream context.