Move beyond generative chatbots. In this lab you'll explore how autonomous AI agents perceive, decide, and act — and what it takes to design them responsibly.
Click each card to reveal the full concept. Remove any one of the five pillars and autonomy collapses.
Tap to expand ↓
Tap to expand ↓
Tap to expand ↓
Tap to expand ↓
Tap to expand ↓
Tap to expand ↓
This is the most important distinction you need to be able to articulate fluently.
Autonomy requires iteration — not a single pass. Click each step to understand what the agent is doing at that moment.
Test your understanding. You need 70% or higher to mark this section complete.
Design your first autonomous agent. Every field maps to a real architectural component — leaving one blank weakens the design.
Autonomy increases both capability and attack surface. Every agent designer must understand these before deploying.
Current LimitationsAgents lose coherence over extended multi-step tasks. They work well for 3–5 steps but degrade significantly beyond that. Design for shorter sub-tasks with human checkpoints.
Context windows are limited. Persistent memory via vector DBs and RAG helps, but is imperfect. Agents can "forget" critical context mid-workflow on complex tasks.
When something goes wrong, agents often compound mistakes by retrying the same failing strategy. They need explicit error-handling logic — not just retry loops.
Malicious instructions embedded in data the agent reads — emails, documents, API responses — can hijack the agent's behavior entirely, turning it against its intended purpose.
An agent with broad data access can inadvertently expose sensitive information through API calls, log entries, or tool outputs. Apply the principle of least privilege — always.
Agents can use legitimate tools in unintended ways — deleting production data while "cleaning up," or sending messages to the wrong recipients at scale.
Synthesize what you've learned. Complete all three prompts thoughtfully — this is what you take out of the lab.