Block 07 / 07
An agent is a loop: perceive, decide, act, observe, under a resource bound.
A completion predicts tokens. An agent takes actions. The durable picture is a loop: perceive the state, decide (sample a thought or a typed call), act (your runtime runs a tool), observe (the result returns to context). Halt on a final channel, a budget, or a guard. ReAct, tool calling, computer use, MCP hosts: same loop, different envelopes.
The model emits a structured request; the host executes it and returns an observation. When the path can be enumerated, the graph belongs to the program and the model fills slots. An open loop is justified when the path cannot be listed. The loop is the invariant; framework names are implementations of it.
Step 1 of 12 · hop 1, perceive
Drag the slider to walk the loop one stage at a time. Watch it halt on its own once the answer is ready.
The loop itself, perceive, decide, act, observe, is a fixed shape. What varies is the decide step: the concrete procedure that turns the current state into the next action. AIMA classes agents as reactive, deliberative, or search-based by exactly that choice. The three families below are that classification applied to LLM agents.
Reactive
One thought, one action, one observation, then repeat, with no separate planning phase and no lookahead: the next action is chosen from the current state alone. Yao et al., 2022 name the pattern; it is the default loop most tool-calling APIs implement directly.
Deliberative
The model drafts a multi-step plan before acting, an executor works through the steps in order, and the loop only returns to planning when a step fails or the observations contradict the plan. Trades ReAct-style responsiveness for fewer, cheaper decide calls on tasks with a stable shape.
Search-based
The decide step expands several candidate next actions instead of committing to one, scores them, and backtracks when a branch dead-ends, as in Tree-of-Thoughts and Language-Agent-Tree-Search. Costs more decide calls per hop; buys recovery from a bad early move without restarting the whole trace.
Paper
The loop written down: thought, act, observation. Read it as a state machine, then map MCP, function calling, and computer use onto the same cycle.
Book
Perceive-act loops, utilities, and environments predate LLMs. The same vocabulary applies to current tool-calling systems.
Course
A current walk through tools and traces. Pair it with AIMA to keep the loop distinct from any one library.
An agent is a loop: perceive, decide, act, observe, under a resource bound. Tool-calling products are envelopes around that loop.
The model emits a typed request. The host runs the tool and returns an observation. Search, calendar, and database access are this cycle, not a hidden ability of the weights.
The observation is a screenshot or accessibility tree. The action is a click or a keystroke. Halt is a budget or a final channel.
A protocol so the same tools can be called from more than one runtime. The loop is unchanged; the envelope is shared.
Implement perceive → decide → act → observe with a max hop count. Two tools: one retrieve, one calculate (or equivalent). Errors come back as observations. Idempotent writes. Emit a JSON trace per hop. One query that must retrieve then calculate. One query that should halt in a single hop. Both are eval cases.