the weather?" — that's tool use, but it's still reactive. True agentic behavior means the system decides on its own that it needs weather data, fetches it, and uses it as part of a larger plan you didn't explicitly outline. ## The four levels of AI autonomy Not all AI systems are equally agentic. Think of it as a ladder: | Level | Name | What it does | Example | |-------|------|-------------|---------| | 1 | **Reactive** | Responds to a prompt, no tools, no memory | Basic ChatGPT conversation | | 2 | **Assistive** | Uses tools when asked, follows instructions step by step | "Search the web for X" — it searches and reports back | | 3 | **Autonomous** | Given a goal, plans and executes multi-step workflows independently | Research agent that reads 50 sources and writes a cited report | | 4 | **Fully Agentic** | Sets sub-goals, delegates to other agents, adapts plans, operates continuously | Sales pipeline agent that runs 24/7, adjusting outreach strategy based on what's working | Most commercial AI today sits at Levels 1-2. The industry is racing toward Levels 3-4. That's the "agentic" shift everyone is talking about. More autonomy means more leverage but also more risk. A Level 1 system can give you a wrong answer. A Level 4 system can take a wrong *action* — and then compound the mistake over dozens of subsequent steps. The higher you go on the autonomy ladder, the more critical guardrails become. ## The agentic loop: plan, execute, observe, reflect, adapt Every agentic system runs a version of this loop. It's the engine that makes autonomous behavior possible: **Plan** — The system receives a high-level goal and breaks it into sub-tasks. "To increase demo bookings, I need to: (1) identify high-intent leads, (2) research each company, (3) personalize outreach, (4) follow up." **Execute** — It takes action: queries the CRM, scrapes a company's website, drafts an email, sends it via API. **Observe** — It checks what happened. Did the email bounce? Did the lead reply? Did the API return an error? **Reflect** — It evaluates whether the action moved it closer to the goal. "My open rate is 12% — below the 25% target. The subject lines may not be working." **Adapt** — It revises the plan based on reflection. "I'll A/B test two new subject line formats on the next 50 leads and compare results." This is the key difference from a basic agent loop (think-act-observe). The agentic loop adds **reflection** and **adaptation** — the system doesn't just do things, it *evaluates* whether those things are working and changes course. That's what makes it agentic rather than merely automated. For each scenario, decide whether the AI is behaving **reactively**, **assistively**, or **agentically**: 1. You ask an AI "Summarize this PDF" and it returns a summary. → ___ 2. You tell an AI "Monitor our social media mentions and draft responses to negative reviews." It checks every hour, drafts responses, and se