A discipline of agentic development
Safety
Agents become truly useful when they are trusted to act. Safety is what makes that trust possible: clear boundaries, careful defaults and people in charge of what matters.
An agent with access to a codebase, a terminal and the internet is powerful, and that power deserves care. The question is not whether to use agents, but how to give them enough room to be helpful without risking what the organisation cannot afford to lose.
I help teams think this through in a calm, practical way: what an agent may do on its own, what needs a human decision, and how to keep secrets, data and production systems out of harm's way.
Three ideas
Least access, most clarity
Give agents exactly the access they need for the task at hand, and make those boundaries visible to everyone.
People decide
Agents propose and people decide, especially where it concerns customers, money, data or production.
Trust that grows
Autonomy is earned step by step. Start small, observe, and widen the boundaries as confidence grows.
In practice
- Clear agreements about what agents may and may not do
- Isolated environments where agents can experiment without risk
- Careful handling of secrets, credentials and personal data
- Keeping an eye on the dependencies and tools agents bring in
Questions to start with
- What could an agent in your team do today that you would rather it did not?
- Where in your process should a human always have the final say?
- How are secrets and sensitive data kept out of reach?