The Cost of Thinking: Architecture, Memory, and Failure Modes

Paper

A Taxonomy of Failure Modes in AI Agents

A large-scale survey of autonomous LLM agent risks. If you're building or deploying, stop guessing about failure modes—it quantifies the real threats to cost and reliability.

Tooling

InterState: Context Window as File System

Local-first memory is the key to reducing latency and cost. This implementation proves you don't need a vector database bill to have persistent, personalized context.

Engineering

LangGraph Launches Production Session Management

Solves the stateful ghost problem. Finally, out-of-the-box setup to checkpoint messy, long-running agent runs without writing ad-hoc caching logic.

Architecture

VeriFact: Fact-Checking the Agent's Scratchpad

Don't trust the agent's notes blindly. This paper introduces a truth-verification layer that stops hallucinated context from poisoning autonomous task chains.

Infrastructure

Reducing Visual Token Cost for Web Agents

Web agents are burning compute on endless renders. This tool cuts the vision processing load significantly by only 'looking' when the text state is ambiguous.

Research

The Dunning-Kruger Effect in LLM Critics

Your agent's internal critique loop is actually making it dumber. This paper shows how

Domain

Automating Database Tuning with LLM Agents

Using LLMs to parse complex SQL execution plans. It's not just automation; it's teaching agents to understand the cause of database performance issues, not just the symptoms.