Embodied, Web, and Software Agents
Agents in the wild
Environment classes
Agentic systems operate in web environments, software repositories, simulators, lab instruments, and physical robots. Each setting changes observability, action latency, safety constraints, and evaluation protocols.
Embodiment matters
In embodied settings, perception is noisy, actions are costly, and errors can be irreversible. This pushes agent design toward robust state estimation, contingency planning, and conservative execution.
Environment characteristics
| Setting | Key challenge | Typical metric |
|---|---|---|
| Web | Dynamic content and prompt injection | Task success |
| Code | Large action space and hidden bugs | Tests passed |
| Robotics | Continuous control and safety | Collision rate |
| Scientific tools | High-cost experiments | Discovery quality |
Operational risk
The closer an agent gets to the physical or financial world, the stronger the case for permission gating, audit logs, and human override.
Which environment property most strongly increases the need for contingency planning?
Irreversible actions make mistakes costly and motivate fallback plans.
Correct answer: Irreversible actions
In software agents, what is a common external verifier?
Code agents often use tests as executable feedback.
Correct answer: Unit tests
Name one reason web agents are difficult to evaluate.
Dynamic content and site-specific behaviors make benchmark stability hard.
Correct answer: The web changes over time
What is one safety mechanism for high-stakes agents?
Other valid answers include sandboxing, permissions, and audit logs.
Correct answer: Human-in-the-loop approval