Agent Framework

The key metric for evaluating agent frameworks has shifted from "single-task success rate" to "stability and recoverability in long-sequence execution".

Current mainstream agent frameworks typically have a large model handle high-level planning and specialized modules handle low-level control, with the two working together to complete long-horizon tasks.

Most agent frameworks do not yet natively provide the ability to isolate a "trusted instruction channel" from an "untrusted data channel."

In most current agent frameworks, the agent inherits trust from its runtime environment, treating accessible documents, web pages, and API responses by default as "readable and executable" context.

Language model agent performance can be significantly improved without updating model weights, an effect achievable through carefully designed agent frameworks.

In both static prediction and dynamic tasks, the framework's pass@1 advantage over stronger interventions (such as model scaling or increased inference compute) may reverse under a larger attempt budget.

When an agent restarts or switches tasks, standard agent frameworks try to rebuild context by replaying 80,000 to 120,000 tokens of chat history.

Agent frameworks usually save long tool outputs as files or in a vector store, and then have the model generate retrieval queries to fetch snippets on demand; the risk of this approach is that the model may retrieve incomplete, imprecise, or non-authoritative snippets, or even skip retrieval entirely.

Over the past eight to ten months, agent frameworks themselves have begun to "self-modify"; Anthropic's talk on workflows demonstrated the ability to dynamically manage sub-agents with topology.

The nondeterminism of tool-call ordering comes from the agent framework's scheduling logic—when the model requests multiple file reads or searches in parallel, completion order is affected by network and filesystem latency, which in turn changes the subsequent context.

Current mainstream agent frameworks generally rely on linear, token-intensive context handling, leading to context dilution, increased latency and soaring costs as project scale grows.

In practice, the interview-style prompt usually requires enabling multi-turn conversation, or running within an agent framework that supports tool calling/follow-up questions.

Most current agent frameworks lean toward using automatic mode (auto) as the default for tool selection, fundamentally because a one-time push combined with automatic mode preserves guiding ability while leaving room to handle unexpected situations.

Token consumption in agentic frameworks is typically several times higher than in simple conversations, because multi-round reasoning, tool call records, and context maintenance are all required.