My Job Is Literally Talking to AI Agents All Day Now

Quick Overview

The presenter details his evolution of AI workflows from simple cloud-based models to complex agentic work using local models within VS Code, highlighting that the primary difference is moving from 'thinking' (model interaction) to 'doing' (agentic execution) which requires managing context and tools more effectively.

Key Points: The presenter currently uses several AI agents in parallel across different workflows, evolving from older cloud-based models to local agentic work within VS Code. The key evolution is the shift from simple model interaction ('thinking') to autonomous agent execution ('doing'), necessitating better context management. Claude Code is demonstrated as a tool that can be used locally within VS Code to manage these agents, offering more control than cloud-only solutions. The presenter set up a local environment using VS Code, the Claude Code extension, Node.js, Python, and an Anthropic API key. The current agentic workflow involves multiple agents running in parallel, each dedicated to a specific task (e.g., website revamp, scraping, market research). A major problem identified is managing context across these parallel agents, as the context window (currently limited to 200k tokens for Claude) can be easily exceeded during complex multi-step tasks. The presentation concludes by previewing a workflow diagram that illustrates how agents communicate to achieve tasks, contrasting this with older automation methods.

Context: The presenter, Steph, outlines the rapid evolution of his workflow for utilizing AI, contrasting older methods that relied on direct interaction with cloud models (like Claude 4.5 via APIs) with his current setup involving local, multi-agent systems running through the Claude Code extension inside Visual Studio Code. He emphasizes that this shift from simple 'thinking' (model prompts) to 'doing' (agentic execution) changes how context and tool access must be managed.

Detailed Analysis

The video details the evolution of the presenter's AI workflow, moving from 'Old school AI' (direct interaction with a cloud model like Claude 4.5 via APIs, often using external tools) to 'AI Automations' (using platforms like Make or n8n to orchestrate API calls to the cloud model), and finally to 'Agentic work' (running agents locally within VS Code via Claude Code). The core difference highlighted is the shift from the AI merely 'thinking' (providing responses) to 'doing' (autonomously executing complex tasks). This agentic setup involves multiple specialized agents running in parallel, each handling a distinct task like website updates, scraping, or market research, all managed through the Claude Code interface running locally on the presenter's computer. The presenter shows his Trello board to illustrate the complexity of managing these parallel tasks. A critical limitation discussed is context window size; while Claude offers 200k tokens, managing complex, multi-step tasks that require feeding large amounts of previous context back into the model remains challenging and is the major problem with current agentic work. He demonstrates the setup requirements (VS Code, Claude Code extension, Node.js, Python, Anthropic API key) and outlines the workflow for setting up a new project, emphasizing that while the local setup offers more power and control, it requires careful management of context to avoid exceeding token limits when agents loop or iterate on tasks.