Boards should govern how AI agents work, not just which tools they buy

In our view, AI agents are ready to take on real work across a company. For boards of established businesses, the important decision is therefore not simply which tool to buy. It is how to let agents act without losing control over decisions, data or delivery. We believe that calls for an agentic harness, built inside the company rather than imposed upon it.

By harness, we mean the organised way agents work alongside people: their roles, rules, permissions, approvals, checks and shared knowledge. It should fit how the company actually works, including where responsibility sits and what counts as acceptable delivery. Our position is straightforward: people decide and agents execute. Delegating a task must not mean delegating accountability.

At DevMojos, we build that harness with the company’s own team. Together, we put specialised agents to work in a few chosen areas and prove the approach on one or two real problems. We train an internal team to operate it and keep improving it. We sell expertise and accompaniment: practical help building the capability, not software, hardware or licences.

The company owns the harness, the agents we build, the rules and the knowledge assembled through the work. It keeps its own accounts and its choice of AI models. We believe ownership should mean practical independence, not merely a contractual assurance. The internal team should understand the system well enough to change it, use several model providers and continue without our help.

Our main promise is speed: a shorter path from request to accepted delivery. We agree a baseline at the start and measure against it. Acceptance matters because producing something quickly is not the same as delivering work that meets the company’s requirements. We do not promise a savings figure before anything has been measured, or agents that never make mistakes. Checks and approvals exist because mistakes remain possible.

Security belongs in the design from the start. Each agent gets only the access its role needs, works in an isolated environment and reaches production only through paths the company authorises, with human approval. Anything irreversible or facing customers requires a named person’s approval. Actions and approvals are logged so the company can audit what each agent did, under whose authority and against which rules.

For us, accountability means named approvers and rules set by people, not responsibility disappearing into an automated process. The company decides where its data may go and which providers are allowed. Its policies come first, and its lawyers decide legal questions. We would never treat an agent’s output, or our own technical judgement, as a substitute for that authority.

Cost control should be equally explicit. Each task should go to the cheapest model that does it well, within agreed budgets and with visible usage. The harness should support several providers rather than make one unavoidable. To manage the pace of change, we establish a routine for testing new models against the company’s own tests. The company decides what to adopt; novelty alone is not a reason to change.

A board should judge the work through a small set of practical questions. Is request-to-accepted-delivery time improving? What is happening to rework and defects? Do agents’ recorded actions match what people approved? Are teams actually using the capability? Can the internal team run it without outside help? Each stage should end with a clear decision on whether to continue, based on evidence rather than enthusiasm.

We believe this is a more useful ambition than buying tools and hoping governance follows. Permission to act should grow only as the company gains confidence in the controls and the results. At DevMojos, we build agentic harnesses with companies, so that AI agents work faster under rules the company owns. The strongest AI programme is the one the board can see, steer and stop.