What Are
AI Agents?

Simple: The agent thinks, acts, looks at the result, and repeats until the job's done.

Deep: The canonical loop is observe → reason → act, formalized as the ReAct pattern (reason + act interleaved in the model's own tokens). At each step the language model emits a thought, selects a tool call, and receives an observation, which is appended to the context window — making the loop effectively a state machine driven by in-context memory. Production systems add a planner (to decompose goals), a critic (to verify outputs), and a controller (to enforce step and budget limits). The model isn't retrained; the scaffold around it is what turns a text generator into an agent.

Simple: It jots things down, so it doesn't forget what it already did.

Deep: Context windows are finite and expensive, so agents use layered memory. Short-term memory is the working context — the transcript of observations and actions. When it overflows, systems apply summarization or compaction. Long-term memory is external: a vector database enabling retrieval-augmented generation, so relevant facts are fetched by semantic similarity rather than held in context. Episodic memory stores past trajectories for reuse. The engineering problem is retrieval precision — pulling the right memory at the right step without flooding the context with noise.

Simple: Small mistakes add up — one wrong turn early can derail the whole task.

Deep: This is compounding error: if each step succeeds with probability p, a chain of n steps succeeds with roughly pⁿ — a 95%-reliable step fails 64% of the time over 20 steps. Mitigations are architectural. Verification loops let the agent check its own work; tool-grounding replaces guessing with deterministic APIs; checkpointing and replanning let it restart from a known-good state; and human-in-the-loop gates catch irreversible actions. Reliability engineering, not raw model capability, is the current bottleneck for agents.

Simple: Today's agents are promising but still make mistakes on long, open-ended jobs.

Deep: Four open problems dominate. Evaluation — there's no rigorous, non-gameable benchmark for long-horizon autonomy; leaderboards saturate quickly. Credit assignment across steps — attributing success or failure in a long trajectory to improve the policy, which is why RLHF-style training doesn't transfer cleanly to multi-step agents. Cost and latency — multi-step, tool-heavy loops are orders of magnitude more expensive per task than a single completion. And safety of action — the difference between a model that is wrong and one that does something wrong is the entire alignment problem for agents. Solving these is the actual research agenda.