A common mistake with AI coding agents is treating a plausible answer or large diff as completed work. An agent can sound certain while misunderstanding the requirement, missing a dependency, skipping a failed check, or changing behavior outside the intended scope. Completion must be defined by evidence.
Prevent this mistake before execution. State the user outcome, constraints, non-goals, allowed area, and acceptance tests. Ask for a plan when the change spans important contracts. Keep work in an isolated branch and restrict credentials and destructive commands. After implementation, require the agent to list changed files, commands run, exact results, unresolved failures, and deviations from the plan. Independently run the relevant checks and inspect behavior, not only code style. Review authentication, data, money, public interfaces, migrations, and rollback with extra care.
Other errors often follow from the same root: oversized prompts, unrestricted autonomy, parallel agents editing the same contract, and merging code nobody understands. They all replace explicit control with confidence.
Use the agent as an execution partner whose work remains inspectable. Start with a small task and make one observable acceptance result the finish line. When a completion claim is easy to verify, you can safely increase scope and autonomy without confusing generated activity with delivered software.
