Ten years ago, much software automation followed scripts and preapproved sequences. Today's coding agents can investigate, choose tools, alter files, delegate subtasks and propose releases. Tomorrow's value will come from letting them do more of that work without asking employees to supervise every step. Businesses will demand the speed, and in many settings they will be right to demand it.
The risk is a slow transfer of authority hidden inside a series of sensible productivity decisions. A company allows an agent to update code. Then it lets the agent run tests, coordinate reviewers, trigger workflows, adjust deployment settings and use whatever credentials appear to make the task easier. Eventually, the company may be very good at recording what its software agents did but less able to show who authorized the important consequences.
That is not primarily a problem of model intelligence. Even highly capable systems cannot invent a legitimate corporate mandate. They may accurately predict what a manager would probably want, but probability is not permission. They may agree with one another, but agreement is not governance. They may possess a token, but a token is not the business reason for using it.
The alternative is not to slow development to human speed. The enterprise can define standing authority for ordinary work: the permissible goals, repositories, data, tools, release environments, time limits and actions that must never be taken automatically. Independent enforcement then evaluates consequential proposed actions against that standing authority before execution. Agents plan and perform work freely within the mandate. When the work requires changing the mandate, accountable people or designated enterprise policies make that decision.
Contra-agents add an indispensable second line of thinking. Their job is to challenge assumptions, expose hidden tradeoffs and ask whether a seemingly reasonable change violates the company's real intent. They are not replacements for enforceable policy; they are independent critics that help reveal when policy needs to intervene.
The choice ahead is not between AI efficiency and enterprise control. It is between autonomy with durable authority and autonomy governed by convenience. The first can scale. The second works until a technically successful agent makes a decision the company never meant to delegate.
Sources
https://owasp.github.io/DevSecOpsGuideline/2-Process/2-2-Develop/2-2-5-AI-Agent-and-MCP-Security/
NEXT STEP