Reasoning and acting together
Early attempts at autonomous AI either reasoned without acting (good plans, no way to execute them) or acted without reasoning (tool calls with no rationale, prone to flailing). The ReAct pattern — short for reasoning plus acting — combines both. At each step the model produces a thought explaining what it intends to do and why, then an action that calls a tool, then receives an observation with the result. The thought makes the model commit to a rationale before acting, which measurably reduces erratic behaviour and makes the agent’s decisions inspectable.
Tool schemas: how the agent learns its abilities
An agent can only act through the tools it is given, and it learns those tools from their schemas. A schema names the tool, describes what it does in plain language, and lists the arguments it accepts along with their types and which are required. The model reads these descriptions and emits a structured call matching one of them. The quality of the descriptions matters enormously: a tool documented as “search” is used clumsily, while “search_company_filings(query, year) — returns official regulatory filings for a UK company” guides the model to call it correctly with the right arguments.
Parsing observations
After every action the agent receives an observation — the tool’s return value. This might be search snippets, a JSON API response, or an error. The agent has to parse that observation back into something it can reason over. This is a frequent failure point: if a tool dumps thousands of lines of raw output, the model loses the thread or exceeds its context window. Well-designed tools return concise, labelled, relevant results, which keeps the agent’s reasoning sharp and its next action well-grounded.
Error recovery in practice
Real workflows fail constantly: searches return nothing, APIs time out, code throws exceptions. A well-built agent treats each failure as just another observation to reason about. Seeing an error, the model can rephrase a query, switch to a different tool, decompose the step further, or report that it is stuck. The mechanisms that make this work are clear error messages (so the model understands what went wrong) and retry limits (so it does not loop forever on the same broken action). Graceful recovery, not flawless first attempts, is what separates a usable agent from a fragile demo.
Putting the loop to work
A complete agentic workflow chains these pieces: a goal goes in, the model thinks, acts via a tool, parses the observation, and repeats — recovering from errors and stopping when the goal is met or a step budget runs out. Understanding the loop is what lets you debug agents effectively. When an agent misbehaves, the answer is almost always in one of four places: weak reasoning prompts, poorly described tools, unparseable observations, or missing error handling. Fix the loop and the agent follows.