Memory Wars, Verification Gaps, and the Agent Reliability Bill Coming Due

Research

Persistent Memory Architectures for Long-Running Agents

This paper digs into why most agent memory is basically a glorified context window dressed up as long term storage. The real unlock is separating episodic memory from semantic knowledge so agents stop re-learning the same lessons every session, which is the difference between a tool and a teammate.

Verification Gaps in Autonomous Code Generation Agents

Turns out agents that write and run their own tests still miss entire classes of logic errors because they're grading their own homework. If you're shipping agentic coding tools to prod, this is the paper that explains why your CI pipeline still needs a human gate for anything that touches money or auth.

Cost of Context Switching in Multi-Agent Task Handoffs

Handoffs between specialized agents burn way more tokens and time than the marketing decks suggest, and this research quantifies it with actual benchmarks instead of vibes. If your architecture has five agents passing a task around like a hot potato, expect your latency and bill to both be worse than a single well-prompted generalist.

Benchmarking Agent Reliability Under Adversarial Tool Outputs

This benchmark feeds agents deliberately malformed or misleading tool responses and measures how badly things fall apart, and the results aren't pretty. Anyone deploying agents with real tool access needs to internalize that most current systems trust their tools way more than they should.

Long-Horizon Planning Failures in Agentic Task Decomposition

Another entry in the growing pile of evidence that agents are decent at breaking tasks into steps but terrible at noticing when the plan itself is wrong halfway through execution. The fix isn't a smarter planner, it's cheap and frequent replanning checkpoints, which most frameworks still don't do by default.

Tools

LangGraph's Approach to Durable Multi-Session State

Durable state keeps showing up because it's the boring infrastructure problem nobody wants to solve but every team building real agent products eventually hits. Worth reading if you're tired of duct-taping Redis and Postgres together to remember what your agent did yesterday.