Unverified50% confidenceFactExact time
The Cognition coding agent team described the phenomenon of 'context rot': models make weaker decisions at longer context lengths.
1
Sources
50%
Confidence
Long-term
Relevance
10/15/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedCount-based prediction can partially identify which inputs are most sensitive to corpus changes, but substantially underestimates the rate at which divergence decreases in neural models.77% similarVerifiedOverly verbose, contradictory, or ambiguous system prompts cause the model's attention to become scattered, resulting in over-thinking or instruction-following failures; streamlining agents.md can improve the determinism of instruction following.76% similarUnverifiedThe study identifies a decision-time leakage problem: the controller peeks at future information that should not be available at the time of decision-making, which systematically overestimates the benefits brought by prediction.75% similarUnverifiedThe next generation of coding agents needs to generate interpretable decision traces—including why a particular solution was chosen, which alternatives were considered, and which parts the agent has lower confidence in—in order to reduce the cognitive reconstruction cost for human reviewers.75% similarUnverified单一智能体面临上下文窗口限制和注意力稀释两大瓶颈，任务复杂度超过单次推理能力时输出质量会显著下降75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1034391API
curl https://kongchang.com/api/v1/knowledge/claims/1034391MCP
get_claim(id=1034391)