LinkedIn publication text
AI Agents vs. Agentic AI: what's the difference?
I compared what the primary sources say about the two terms. The uncomfortable conclusion: there is no consensus definition.
🔎 What I found
- Only the taxonomy by Sapkota et al. (2025) treats the terms as distinct categories: agent = module for a specific task; agentic AI = coordinated specialized agents with persistent memory.
- Anthropic separates workflows (predefined paths) from agents (the LLM drives the process).
- OpenAI defines an agent as a system that acts independently and treats “agenticness” as a degree, not a category.
- Google bets on layered architecture and the A2A protocol; MCP and AGENTS.md moved to the Agentic AI Foundation (Linux Foundation).
⚠️ The flip side
- Multi-agent systems can use about 15× the tokens of a chat (Anthropic).
- The MAST study mapped 14 failure modes in 3 categories.
- OWASP published a Top 10 of risks for agentic applications.
- Platforms shift fast: OpenAI shut down Agent Builder and Evals in Jun 2026.
✅ What to do
- Start with the simplest solution.
- Test end to end and by subtask before granting autonomy.
- Keep people in high-risk actions.
- State the definition you adopt.
In the carousel: 29 slides, glossary, a 2022–2026 timeline, comparison tables and 15 references graded by verification level (A, B or C). Where I only found a secondary source, I marked it with †.
Which definition does your team use?
© Aridio Silva @aridiosilva · October 2026