LLM Agent

In goal mode, Agent behavior can be quite aggressive, and it is not recommended to enable it when the goal's achievability is uncertain or when cost is a concern.

Prompt injection is a novel and unique attack vector targeting operational agents; malicious web pages can inject instructions into an agent via visual elements or hidden text, and how to distinguish real instructions from forged ones remains an open problem that has not been fully solved.

现代LLM-based Agent的核心突破发生在2023年前后，随着GPT-4、Claude等大模型具备了稳定的指令跟随和工具调用能力

A typical Agent execution loop includes perception, reasoning, action, and observation, forming the ReAct (Reasoning + Acting) pattern

When the number of tools an Agent can call grows from single digits to dozens or even hundreds, cramming all tool descriptions into the context occupies a large portion of the context window and reduces the model's selection accuracy.

The paper names the safety failure mode arising from skill composition the Conflict of Skills (CoS), referring to the possibility that artifacts or evaluation signals generated for one purpose may, when reused downstream, fail to preserve the constraints or authorizations required by the new purpose.

An LLM Agent is composed of a model and an execution harness, and its benchmark performance reflects the joint behavior of the two.

The agent compiled each finding into an executable test, and the test suite scans the entire corpus in only 0.7 seconds of wall-clock time without a GPU.

Each of the three agent runs cost $2.21 to $2.84, and together they found seven issues across four categories of phenomena.

LLM agents can serve as an effective corpus debugging tool, but should not fully replace static baseline methods.

The best static sampling baseline validated more selectors than any agent run, at less than half the cost of an agent run.

The agent's core contributions lie in accurate prevalence estimation, precision estimates checked against sampled documents, explanations of why the pipeline misses each slice, and reusable tests.

The study used an LLM agent as a corpus debugger to audit the Slovak portion of FineWeb2 (26.5 million documents).

In the LLM Agent context, prefetch mode refers to an external orchestration layer calling data-source APIs in advance before the model begins reasoning, then formatting the results and injecting them into the prompt as context.

Current LLM Agents typically rely on implicit inference from the match record within the context window and lack an explicit belief-state maintenance mechanism, making them prone to forgetting early clues or overfitting to recent turns in long games.

The core idea of such LLM Agent Pokémon battle projects is to convert the battle interface into a text state that the LLM can understand; the Agent reads the state and outputs action commands, which a script then sends back to the battle platform for execution.

Having models play games is nothing new, but Pokémon battles require models to possess three capabilities at once: rule understanding, opponent modeling, and risk trade-off.

Ignoring the distributed nature of agent retries leads to three types of consequences: data corruption (duplicate transactions, lost updates, inconsistent state), resource exhaustion (infinite loops consuming API tokens and compute), and unpredictable behavior.

All three types of agent retry failures above (non-idempotent operations, state drift, and cascading failures) occur at the system layer rather than the model layer.

LLM Agents are stateful, non-deterministic processes that interact with external systems and cannot simply be treated as HTTP clients for retries.