The agentic coding loop
Move from a request to an evidenced change in four deliberate steps.
Inspect: establish what already exists
An agentic coding tool can read files, edit code and run checks. Start by asking it to inspect the project and explain the current behavior with file references. For a teaching board, ask where sessions are loaded, where filtering happens and how assignments are saved. This prevents a convincing plan built on an imaginary architecture.
Plan: agree the smallest coherent change
Describe the desired behavior and constraints. For example: add a “Needs facilitator” filter to the existing board, reuse its data model and show a useful empty state. A plan should name the code boundary it will touch and how the result will be verified. You do not need an essay or a new architecture for a small change.
Build: keep the change inspectable
Ask the agent to implement a bounded piece. If the patch expands into login, calendar synchronization and email reminders, bring it back to the acceptance criteria. Small changes are easier to understand, test and undo. Parallel agents help only when their work can be separated cleanly and then integrated.
Verify: close the loop
Check the actual behavior, read the diff and inspect the test evidence. Ask “What could still be wrong?” A successful command proves only what that command checks. Then explain the change in your own words: input, transformation, output and one failure case. This turns delegation into learning.
Set a budget and a stopping condition
Name a bounded outcome, the checks that establish it and the conditions that require a handover. For an automated loop, enforce step, time and spend limits before the next operation. If a check fails, record the failure and choose a justified next step; do not let repeated attempts quietly become an unbounded job.