Skip to content

AI Tool Selection in the Agentic Era

Category: guide Last updated: 2026-04-07 Status: complete

Summary

A practical framework for selecting which AI tools to use, based on Ethan Mollick's February 2026 guide. The key insight is that AI now consists of three separable layers — Models, Apps, and Harnesses — and understanding each layer is necessary to get meaningful results. The shift from conversational AI to agentic AI (systems that autonomously complete multi-step tasks) represents the most significant change since ChatGPT's launch.

Details

The Three-Layer Framework

Understanding AI tools requires distinguishing between three components:

Models are the underlying AI brains — the core intelligence that determines reasoning ability, writing quality, coding capability, and image processing. As of early 2026, the three leading frontier models are:

Model Provider Notable Strength
GPT-5.2 / GPT-5.3 Thinking OpenAI Complex analytical and statistical work; Extended and Heavy variants for increasing difficulty
Claude Opus 4.6 Anthropic Writing quality; extended thinking feature; limited image generation
Gemini 3 Pro Google Competitive with other frontiers; Deep Think variant for hardest problems; weaker harness ecosystem

Apps are the products users interact with directly — primarily the main chat interfaces (chatgpt.com, claude.ai, gemini.google.com) plus specialized tools like concepts/claude-code and OpenAI Codex.

Harnesses are the environments that give AI tools, autonomous action, and multi-step task capability. The same underlying model produces dramatically different results depending on its harness. This is why the "which AI is best?" question is increasingly inadequate — the harness often matters more than the model.

Why Model Selection Within Apps Matters

A critical and non-obvious finding: default model selections within apps are often not the best available option. Apps frequently route users to weaker variants by default. Manually selecting advanced models (e.g., choosing GPT-5.2 Thinking Heavy instead of the default) yields substantially better results.

Action required: When opening any AI app, explicitly select the most capable model rather than accepting the pre-selected default.

Specialized Applications Worth Knowing

For Technical Users

concepts/claude-code — Anthropic's agentic development environment. Provides access to a virtual computer, web browser, and code terminal. Enables autonomous software development from research through testing. Best for: software development, complex automation, data processing.

OpenAI Codex — Similar to Claude Code; provides terminal and codebase access. Best for: software development, code-heavy tasks.

For Non-Technical Users

Claude Cowork — Desktop application from Anthropic for non-technical work. Operates in an isolated virtual machine for security. Organizes documents, extracts data, drafts summaries. Best for: document management, data extraction, writing tasks.

Claude for Excel / Claude for PowerPoint — Spreadsheet and presentation extensions. The Excel integration is described as particularly impactful for analysts. Best for: spreadsheet automation, presentation creation.

OpenClaw — Open-source local agent with web browsing, file management, and email capabilities. Free but carries significant security risks due to broad computer access. Best for: users who understand the security tradeoffs and want local control.

For Research and Knowledge Work

NotebookLM (Google) — Accepts documents, videos, websites, and files as input. Generates interactive knowledge bases, slideshows, mind maps, and AI-generated podcasts that discuss source material. Has a free tier. Best for: research synthesis, literature review, learning from documents.

Recommendations by User Level

For beginners: 1. Pick one system (ChatGPT, Claude, or Gemini) 2. Subscribe at $20/month 3. Always manually select the most advanced available model 4. Learn its capabilities before branching out

For proficient users: 1. Use NotebookLM for research synthesis (free tier available) 2. Use Claude Code or Cowork for advanced automation 3. Try Excel/PowerPoint plugins for office productivity 4. Experiment with OpenAI Codex for code-heavy tasks

The Agentic Shift

The fundamental change described by Mollick: AI has moved from conversation to action. Rather than chatting about a task, you can now assign it — the AI uses tools, makes decisions, and returns with completed work.

This shifts the relevant questions from "which AI writes best?" to "which AI-plus-harness combination accomplishes this task best?" The harness determines what tools the model can use and how autonomously it can act.

See also concepts/agentic-engineering for the engineering-focused perspective on this shift, and concepts/ai-inflection-point for the November 2025 threshold where agentic reliability crossed into production-viable territory.

Key Claims & Data Points

  • The three leading models as of early 2026: GPT-5.2/5.3, Claude Opus 4.6, Gemini 3 Pro — [source: a-guide-to-which-ai-to-use-in-the-agentic-era]
  • Default model selections in apps are often weaker variants; manual selection substantially improves results — [source: a-guide-to-which-ai-to-use-in-the-agentic-era]
  • Gemini 3 Pro has a weaker harness/application ecosystem despite a competitive base model — [source: a-guide-to-which-ai-to-use-in-the-agentic-era]
  • The agentic shift represents "the most significant change since ChatGPT's launch" — [source: a-guide-to-which-ai-to-use-in-the-agentic-era]
  • NotebookLM (Google) has a free tier and generates AI podcasts from source material — [source: a-guide-to-which-ai-to-use-in-the-agentic-era]

Open Questions

  • How rapidly are these model rankings changing? The article is from Feb 2026 and the landscape shifts monthly.
  • What is the precise definition and scope of "Claude Cowork" — is it a separate product or a mode within Claude?
  • How does NotebookLM's knowledge base quality compare to alternatives like the concepts/llm-knowledge-base pattern?
  • Does "harness" in Mollick's framework map directly to "harness engineering" in the OpenAI Codex team's usage, or are these different concepts?

Sources