What was broken
High-value compression-run setup depended on a lengthy senior-engineering workflow. It was reliable. It was also the throughput constraint: higher volume meant more expert hours the team did not have.
Product decision
Treat the bottleneck as an operations product, not a chatbot. The agent receives authorized live context (run state, configurations, client data) and acts only through typed, bounded surfaces. Operators stay in the loop.
Flow
- 01Before3.5h expert setup
Senior engineering was the throughput bottleneck.
- 02DecisionBounded agent in the workflow
Live context, typed actions, operator control.
- 03After8 minutes, 3× volume
Operations absorbed the load without added headcount.
System pattern
- Assembles only the context needed for the current run
- Uses controlled tool actions rather than open-ended system access
- Keeps operator visibility and intervention in the execution loop
- Measures setup speed, throughput, and configuration quality in production
Role & tags
Lead Product Manager. Owned 0→1 strategy and launch: framed the workflow, defined the agent boundary and success metrics, designed controls and rollout, and tied the product to operating leverage.
- Bounded autonomy
- Human visibility
- Live context
- Production quality metric
Impact
Setup fell from 3.5 hours to 8 minutes. Run volume scaled 3× at constant headcount. Zero AI-initiated configuration errors in the first six months.
1-6 beats · j / k · ← → adjacent · y link · b brief