Cache Me If You Can: Designing Databases for Swarms of AI Agents
LLM agents do not merely query databases; they speculate against them, and that changes what modern data platforms must optimise.
Cognaptus Insights
Industry thoughts, trends, and commentary.
Control Plane, Not Pain: How Agentic OS Turns Linux Scheduling into a Semantic Service
A mechanism-first reading of SchedCP, an agentic control plane that turns Linux scheduler tuning from kernel craft into workload-aware infrastructure optimisation.
From Prompts to Policies: The Agentic RL Playbook
A mechanism-first guide to why Agentic RL reframes LLMs as trainable policies operating across tools, memory, environments, and long-horizon business workflows.
Judgment Day for RAG: How L‑MARS Cuts Legal Hallucinations by Design
L-MARS shows that retrieval helps legal AI when the answer depends on current authority, but adds little when the task is mainly legal reasoning.
Map Before You Train: Data Cartography to Defuse LLM Memorization
A practical reading of GenDataCarto: how per-sample training dynamics can expose memorization hotspots before they become privacy, contamination, or governance failures.
Brains Meet Brains: When LLMs Sit on Top of Supply Chain Optimizers
A mechanism-first reading of how LLMs can make supply chain optimizers usable without pretending to replace the optimizer.
Dial M—for Markets: Brain‑Scanning and Steering LLMs for Finance
A mechanism-first reading of how sparse autoencoders can expose and steer the concepts an LLM uses when processing financial news.
Mask, Don’t Muse: When Simple Memory Beats Fancy Summaries
A practical reading of why simple observation masking can beat expensive LLM summaries for software-engineering agents.
Numbers Need Narration: Making LLMs Do Reasoning‑Intensive Regression
A mechanism-first look at why LLM scoring systems need both reasoning and calibration when text must become a precise number.
Patience Is Profit: Can LLM Agents Stabilize DePIN’s Token Rails?
A mechanism-first reading of EconAgentic, a DePIN market simulation showing how token incentives, node-provider patience, and LLM-style decision rules may affect network inclusion, stability, and market value.