Why Mastering AI Agents Is the Most Critical Skill Right Now

作者：ybq
链接：https://zhuanlan.zhihu.com/p/2025676117307138196

Why Code‑Centric Agents Matter

Code acts as the interface that lets large language models (LLMs) manipulate real‑world artifacts such as files, APIs, and compute resources. Modern agents (e.g., Claude Code, Codex) can autonomously browse the internet, modify source code, run experiments, and generate reports, achieving productivity that rivals or exceeds a human specialist.

Core Capability: The Environment

LLM development has progressed through three functional stages:

Chat Model – human‑level IQ/EQ for conversational tasks.

Reasoning Model – graduate‑level knowledge and logical inference.

Agent Model – workplace‑level autonomy that can discover problems and act within a defined environment .

The environment supplies three essential ingredients:

Permissions (read/write/search, execution rights, network access).

Tooling (compilers, debuggers, package managers, sandboxed runtimes).

Reward Signal – stable, fair feedback that tells the agent whether an action was successful.

When these are rich and reliable, the agent’s exploration space expands, leading to higher subjective agency and better tool‑selection timing.

Training Paradigms

Two dominant research tracks aim to improve code agents:

Supervised Fine‑Tuning (SFT) – “hacking Claude cot is all you need”. The model learns by imitating expert demonstrations (e.g., best‑practice code edits, data‑analysis pipelines).

Reinforcement Learning (RL) – “scaling RL is all you need”. The model learns through trial‑and‑error, receiving rewards from the environment for successful executions (e.g., passing tests, reducing runtime errors).

Both approaches share the same ultimate goal: exploring the environment effectively. SFT provides proven exploration patterns; RL refines those patterns by letting the model discover new strategies.

Comparative Trade‑offs

SFT yields high short‑term efficiency because the agent copies proven workflows, but it may struggle with out‑of‑distribution tasks.

RL promotes continual growth and adaptability, yet it requires substantially more compute and often exhibits lower token efficiency during training.

The decisive factor is not the algorithmic choice but the richness of the environment: broader permissions, more diverse tools, and consistent reward signals raise the ceiling of what any agent can achieve.

Illustrative Agent Capabilities

Search the internet for a set of STEM problems that a baseline model cannot solve; in one night the agent can generate solutions surpassing a PhD intern’s weekly output.

Given a GPU‑enabled sandbox, the agent can add a new feature to an RL codebase, run 20 training steps without errors, analyse logs, iteratively fix bugs, and complete in a single evening work that would normally take two weeks.

For routine data‑analysis, a single prompt (“show me this dataset”) triggers the agent to locate the file, infer relevant columns, perform statistical processing, and return a formatted report—no explicit field‑by‑field instructions required.

Practical Adoption Guidance

To harness agents effectively, practitioners should:

Provision a sandbox with full read/write/search access to code repositories, data stores, and external APIs.

Integrate tool wrappers (e.g., git, docker, python -m pytest) so the agent can invoke them via LLM‑driven commands.

Define clear reward metrics—such as test‑suite pass rate, execution time reduction, or data‑quality scores—to guide RL fine‑tuning.

Combine SFT (to bootstrap competence) with periodic RL cycles (to refine and expand capabilities).

When multiple agents are deployed, they can cross‑validate each other’s outputs, further improving reliability.

Strategic Outlook

Agents that can read an entire codebase in minutes and operate 24/7 dramatically increase throughput. The current performance gap between leading proprietary agents (Claude Code, Codex) and domestic alternatives is comparable to the impact of earlier breakthroughs such as DeepSeek‑R1. Closing this gap will likely require releasing models that match at least 90 % of the user‑experience capabilities of Claude Code.

In the longer term, LLM agents are viewed as a stepping stone toward artificial general intelligence (AGI); future AGI breakthroughs are expected to emerge from tight collaboration between agents and human developers rather than from isolated model scaling.