flowchart LR A[Goal and constraints] --> B[Model proposes next action] B --> D[Harness checks permission] D --> E[Allowed tool reads, edits, or runs] E --> F[Observed result] F --> B F --> G[Human reviews artefact]
7 From chat to agent
A coding agent connects a language model to tools that can inspect files, make changes, and run code. The harness is the software that carries messages between the model and those tools, applies permission checks, and shows you what happened. Some chat interfaces also use tools; the useful distinction is what the system can act on and how you supervise those actions.
7.1 The loop
The cycle is observe, decide, act. The model receives the task and available evidence, proposes a next step, and receives the tool’s result. An error can lead to another inspection or repair. A denied action should lead to a stop or clarification when proceeding would exceed the agreed scope. The loop does not need a new human prompt for every step.
For example, a synthetic questionnaire contains a repeated item. The agent might read the column names, compare the two rows, write a corrected copy, and run a check. Each result should narrow the next decision. Simply deleting everything with a repeated item name could also remove legitimate versions from different waves.
The system’s state is what has accumulated so far: conversation, files, settings, and running processes. A message saying “I will fix it” changes the conversation; a successful edit changes a file. “All checks passed” is still a claim until you inspect the command result and the actual output. An agent can stop early, check the wrong file, or describe work it never performed.
Autonomy means how much the system may do before returning to you. Define that boundary through allowed folders, permitted actions, checks, and a stop condition. In Lab 1, watch the reads and approvals, inspect the corrected questionnaire, and compare the agent’s account with the checker result.
A CLI (command-line interface) exposes this work in a terminal; a GUI (graphical user interface) uses windows, buttons, and visual controls. Appearance alone does not establish safety or autonomy. Compare available tools and boundaries.
TipTry it now (5 minutes)
Ask an agent: “List the files you would inspect for this task. Do not read or edit anything.” Then ask which statements are observations and which are guesses.
7.2 Further reading
- Building effective agents — Distinguishes fixed workflows from agents and shows why simple tool loops and stop conditions are often enough.
- ReAct — The influential reason–action–observation pattern behind the loop on this page; its early results are useful context, not a guarantee for current agents.
- Language Models Don’t Always Say What They Think — Shows that a plausible explanation can omit the cue that actually changed an answer, which is why we inspect actions and outputs.