Agents
The Assembly Line: observe, decide, act, verify
What is an AI agent?
Lab honesty
Agent or ordinary software?
Prefer deterministic code for stable rules, exact calculations, and known sequences. An agent becomes useful when the model must choose among tools or adapt the next step from an observation. Even then, deterministic code should enforce schemas, permissions, budgets, approvals, and the final stopping condition.
Agent loop experiment
Choose who controls the path, set hard budgets, and optionally make the first simulated tool call fail. Every run is deterministic: the same task and settings produce the same trace.
Pick a task
Observable state
Run a task, then pause or reveal one step at a time to inspect the loop.
What can go wrong
- Infinite loops: the agent keeps calling tools without making progress
- Hallucinated tool calls: AI invents a tool that doesn't exist
- Token budget overruns: complex tasks can consume thousands of tokens across many steps
- Getting stuck: the agent may retry the same failed approach repeatedly
- Prompt injection: untrusted tool or document content may try to redirect the model away from the user's goal
- Excess authority: a plausible but wrong action can cause harm when tools have broader permissions than the task requires
Key Insight
See also: Unhobbling
Check your understanding
1/4What makes an agent different from a single AI call?
Sources, scope & review statusVerified Jul 29, 2026
Claims this lesson makes
- Stable conceptAgent systems iterate over observations, decisions, actions, and stopping conditions.
Scope: Displayed plans are observable summaries, not a claim to expose a model's private chain of thought.
- Stable conceptAn agent run needs explicit exit conditions, bounded tools and permissions, and a handoff path when the system cannot complete the workflow safely.
- Stable conceptAdding more agents is not automatically better; orchestration adds coordination, evaluation, latency, and maintenance costs.
- Stable conceptExternal content can carry indirect prompt injection that redirects an agent, so untrusted observations must not be treated as authorization for high-impact tool actions.
Primary reading
Next content review: 2027-07-29. Source links establish the lesson's claims; interactive numbers remain labeled simulations unless explicitly identified as measured data.