The First Decision I'd Hand to an Agent
My last article was about migrating my commissioning tracker to AWS, with AI helping design and deploy it. That raises the obvious follow-up: once the app lives on AWS, what should the AI actually do inside it? Not everything. One specific thing. Every commissioning project has a repeated judgment call that eats my mornings, and it's the first decision I'd hand to an agent: issue triage.
The decision
On a live liquid cooling project, issues arrive constantly. A Modbus timeout on a coolant distribution unit. A network switch needing a firmware patch. A client request to rearrange an HMI alarm screen. Someone has to read each one and decide four things: priority, discipline, location, and owner.
That call is repeated, judgment-heavy, and mostly pattern-based. A leak detection alarm outranks almost everything. A cosmetic HMI tweak can wait. An experienced engineer makes these calls in seconds, but seconds multiplied by hundreds of issues across a project is real time, and it's the same shape of decision every single day.
What the agent would need to get right
Trust is the whole game here, so my requirements are strict:
- Real equipment context. The agent must know a CDU from a network switch and what a BACnet mapping error implies about the BMS integration. Generic text classification is not enough.
- A conservative bias. When unsure, escalate to me. Never guess on safety-related issues. I'd rather review ten extra items than have one critical leak alarm sit in a queue.
- Shown reasoning. One line explaining each assignment, so I can spot-check in seconds instead of re-reading every issue.
- No silent learning curve. It starts as a suggestion engine. It earns autonomy by being right, logged, over weeks.
How I'd build it
This is where being on AWS pays off from the migration. The issue data already lives in DynamoDB, so an agent built with the Strands Agents SDK on Amazon Bedrock can read new issues as they land, pull project context, and write back a suggested triage with its reasoning. Lambda triggers it, everything is logged, and the whole thing costs pennies because it only runs when an issue arrives. Bedrock's model choice also matters for client confidentiality: the data never leaves the boundaries I already set.
How I'd know it's working
One metric: my edit rate. If I stop correcting its priority and owner assignments, it's working. If my morning review drops from thirty minutes to five, it's really working.
The honest caveat from my earlier articles still stands. The agent triages; it never closes a safety issue, never talks to a client, and never touches a control system. Judgment about consequences stays with a person. Sorting the inbox is delegation. Owning the outcome is not.
I'll report back with real numbers once it's running on live project data. Current progress is at my engineering website www.terrylaws.com .
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article