AI Agents

Agent frameworks are hard to fuzz: defects surface as ordinary exceptions or silent semantic failures rather than crashes, and Pydantic type governance makes naive fuzzers emit mostly invalid inputs. LogicHunter fuses formal type constraints with real-world usage patterns to synthesize inputs that are valid by construction yet semantically extreme, then resolves oracle ambiguity with a ReAct-based agent that reads docs, walks source, and inspects runtime state. It finds real bugs across LangChain, LlamaIndex, and CrewAI—a concrete win for testing infrastructure you already depend on.

MCP renders tool metadata once in an approval dialog, then injects it verbatim into the model every turn—and nothing requires the rendered view and the delivered bytes to match. The author isolates that gap and shows Unicode's TAG block (U+E0000–U+E007F) has no glyph in mainstream terminals, chats, or IDEs, so a payload written in it is invisible to reviewers but survives byte-for-byte into the tokenizer. Across 8 techniques on a real JSON-RPC/stdio harness, all reach the model, 4 evade a string sanitizer, and only the TAG-block encoding is fully invisible.

NapMem treats long-term user memory as a structured action space rather than a passive retrieval channel. It organizes history into a multi-granularity pyramid—raw conversations, typed records, topic tracks, profiles—linked by provenance, and exposes each level as a tool the agent learns to query via reinforcement learning. On PersonaMem-v2, LongMemEval, and LoCoMo it is competitive on memory-intensive tasks while largely preserving general reasoning and tool-use on non-memory tasks. The ablations over navigation, granularity, and RL training are the part worth reading.

AgentTether does runtime repair without touching the underlying agent or environment. It abstracts each run into Transition Units, links them through a dependency-aware Critical Transition Graph, and localizes failure-critical subtrajectories by combining an offline normal-behavior model with a run-local detector—then converts the cause into behavior-scoped guidance backed by cross-iteration repair memory. On 261 tau-bench tasks across three domains with Qwen3.7-max, it repairs 59% of the hardest Banking failures and transfers cross-model to GPT-5.4. Diagnosis beats blind retry; the localization step is what makes the guidance grounded.

This challenges the assumption that in-loop retrieval is too slow. The authors argue latency is a property of where the store lives, not the pattern: an in-process store answers in ~100us, three orders below a network round trip, so reading and writing memory every step stops costing anything. Holding a fixed latency budget and varying only store speed, redundant actions rise monotonically with latency—0 of 12 in-process versus 7.2 of 12 at 110ms (p=0.0079). Recall climbs from 0/5 to 3.6-4.8/5, though an instructed restate baseline also solves it.