Pillar 2: Building agents
How do you bring an AI agent into production safely?
Definition
An agent goes to production in five steps: design, build the core, shadow running, limited autonomy and scaling up. Between each step sits a measurable gate, so an agent earns its autonomy with results.
- 1DesignGate 1: Value and risk agreed
- 2Build the coreGate 2: Passes the golden set
- 3Shadow runningGate 3: Matches the team's decisions
- 4Limited autonomyGate 4: Stable in production
- 5Scale up
Explanation
- Design: the process mapped, a draft agent passport and an agreed measure of success.
- Build the core: the smallest agent that already delivers value, inside the agentic harness.
- Shadow running: the agent runs alongside on real cases without acting itself yet. You compare its proposals with what the team decides.
- Limited autonomy: the agent acts on its own for low-risk cases, people handle the rest.
- Scale up: more cases, more sources, then the next department.
That is how an agent earns its autonomy with evidence.
What it looks like in practice
A finance department lets an agent run alongside for a month on invoice matching. If its proposals match the team's decisions in the vast majority of cases, it may book standard cases itself.
FAQ
Frequently asked questions
The agent runs alongside on real cases without acting itself. You compare its proposals with the decisions people make.
