Autonomous Agents

Definition

Agents are "LLMs using tools based on environmental feedback in a loop." They gain "ground truth" from tool results / the environment at each step to assess progress.

Key points

• Lifecycle: begin with a command or interactive discussion from a human; once the task is clear, agents "operate independently, potentially returning to the human for further information or judgement." They can "pause for human feedback at checkpoints or when encountering blockers."
• Need guardrails and are best run in sandboxed environments with stopping conditions (e.g., max iterations) to control cost/errors.
• When to use agents over workflows: "open-ended problems where it's difficult or impossible to predict the required number of steps, and where you can't hardcode a fixed path."
• Three implementation principles: (1) maintain simplicity; (2) prioritize transparency by showing planning steps; (3) carefully craft the agent-computer interface (ACI) through tool documentation and testing.
• ACI / poka-yoke your tools — error-proof tool arguments so mistakes are harder to make; invest in tool documentation as much as in prompts.

Why it matters for the exam

• "Agent = LLM using tools in a loop" and the ACI / poka-yoke terminology are testable, as is the rule for choosing agents over workflows (unpredictable step count, no hardcodable path).

Common gotchas

• Start simple; only add agentic complexity when it demonstrably improves outcomes — agents trade latency and cost for autonomy.
• Frameworks can obscure the underlying prompts/responses and tempt added complexity; consider starting with the LLM API directly.

See also

• Workflows vs Agents
• Augmented LLM
• Orchestrator-Workers Workflow
• Evaluator-Optimizer Workflow
• Agentic loop
• Hub-and-spoke orchestration

Sources

• blog-building-effective-agents