我的「破產版」賈維斯：調教 AI Agent 的血淚史

2026-03-06 · Source: AI on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Intermediate, short

Summary

A developer recounts the challenging process of training an AI agent, dubbed "Budget Jarvis," to automate web-based data retrieval using Claude, Playwright, and Brave API. Initial attempts led to excessive token consumption and system crashes due to bloated web structures and the AI's tendency to "memorize" irrelevant page elements. The agent also exhibited "confirmation OCD," performing redundant steps for simple form filling, and a "lack of judgment," opting for browser-based navigation over more efficient API calls for straightforward queries. Through iterative prompting and system adjustments, including a sliding window approach for context management and explicit instructions on planning and cost awareness, the developer significantly improved the agent's efficiency and accuracy.

Key takeaway

For AI Architects designing autonomous agents, recognize that simply providing powerful tools like large language models and browser automation is insufficient. You should prioritize teaching your agents judgment and cost awareness through careful prompt engineering, rather than solely focusing on increasing context windows. Explicitly instruct agents to plan multi-step actions and to weigh the efficiency of different tools (e.g., direct API calls vs. browser navigation) to prevent excessive resource consumption and improve operational speed.

Key insights

Effective AI agent training requires teaching judgment and cost awareness, not just providing powerful tools.

Principles

• Streamline processes over increasing context window size.
• Introduce cost awareness into AI decision logic.
• Plan actions before execution for efficiency.

Method

Implement a sliding window for context, compressing web page elements by ignoring non-essential sections. Instruct the AI to plan all steps before acting and to weigh tool usage based on "cost awareness" (e.g., API vs. browser).

In practice

• Use a sliding window for AI agent context.
• Prompt agents to "plan first, then act."
• Teach AI agents "cost awareness" for tool selection.

Topics

• AI Agent Development
• Web Automation
• Large Language Models
• Prompt Engineering
• Token Management

Best for: AI Architect, AI Engineer, Machine Learning Engineer, Prompt Engineer

Related on AIssential

• I Thought One Prompt Could Do Everything… I Was Wrong — Multi-agent AI systems, mimicking human teamwork, overcome single-model "context overload" by distributing tasks.
• I Kept Hearing About AI Agents Everywhere — So I Finally Sat Down and Learned What They Are — AI agents are goal-pursuing systems that reason, use tools, and operate in loops, fundamentally changing software interaction.
• Prompt Engineering for Agentic AI — Agentic AI prompting requires architectural context engineering, not just better phrasing, to ensure reliable multi-step behavior.
• QCon London AI Coding State of the Game: More Capable, More Expensive, More Dangerous Coding Agents — AI coding agents are more capable but introduce significant security risks and increased operational costs.
• Why Notion’s biggest AI breakthrough came from simplifying everything — Simplifying AI interactions with human-readable formats and minimal abstraction dramatically improves model performance and user adoption.
• Level-up your AI Agent with Skills Engineering — Skills engineering explicitly guides AI agent decision-making through detailed instructions, improving reliability and performance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI on Medium.