Agentic Everything

  • Author/Source: Teddy Svoronos (Harvard Kennedy School), via Substack

  • Original: https://tedsvo.substack.com/p/agentic-everything

  • Key Ideas

  • An agent is an LLM that runs tools in a loop to achieve a goal (Simon Willison's definition). Deep Research runs web searches in a loop to write a report; Claude Code executes code in a loop to build software; Claude Cowork runs local commands in a loop for any task.
  • November 2025 marked the start of a "categorically different thing" --- agents now work with local files, calendars, and metadata, mapping out plans, spawning sub-agents, and making complex decisions autonomously.
  • Subject-matter expertise is more useful than before, not less: directing agents effectively requires strong mental models of what should be happening, and the ability to spot when things go wrong.
  • Getting good at agentic tools takes dedicated practice --- it is a separate skill from domain expertise.
  • The author argues that subject expertise and AI skills should be taught as separate, discrete units rather than interleaved, because mixing them risks dulling the development of both.

  • The biggest pedagogical shift: we must prepare students not just for AI doing their assignments, but for AI being able to do the jobs we train students for.

  • Summary

Svoronos reflects on how the latest agentic AI models (Claude Opus 4.5/4.6, ChatGPT 5.2, Claude Code, Claude Cowork) represent a qualitative leap from earlier chatbot-era tools. Rather than assisting with individual steps, these agents can now own entire tasks --- planning, iterating, self-testing, and delegating to sub-agents --- often running autonomously for 25-30 minutes at a stretch. Svoronos describes feeling, for the first time, like he is "supervising a fleet of researchers."

The post pivots to implications for teaching. The core tension is that while domain expertise remains essential for directing agents, learning to use agents well is itself a substantial, practice-intensive skill. Svoronos concludes that courses should separate the two: maintain (and update) rigorous subject-matter instruction while reserving dedicated time for hands-on agentic tool practice. He leans toward redesigning his statistics course to focus more on judgment and evaluation, abstracting some traditional building blocks, and adding a multi-week module on directing agentic tools. He acknowledges significant risks: not all students will have access to top tools, capability varies across providers, and the landscape may shift again soon.

  • Relevance to Economics Research

This article directly addresses how social science faculty (especially those teaching quantitative methods) should adapt pedagogy in the agentic era. It frames the key challenge for economics researchers: domain expertise becomes the critical differentiator when AI can execute technical work, but researchers must also invest in learning the agentic toolchain. The discussion of delegation, quality control, and "supervising" AI resonates with how economists might restructure their own research workflows.