OpenClaw setup
Dedicated Mac, controlled tools, its own logins, narrower worker agents.
Guest lecture · agentic systems
From prompts
From prompts
to operating systems.
A short talk about moving from one-off prompting to systems that can carry context, route work, and support human judgment in real environments.
Quick background
Why I started building systems.
I’m Matt Yilmaz, a UT Government senior. This class pushed me from one-off prompting toward workflows that could actually carry work.
I was not trying to build a sci-fi agent. I was building an operating layer.
Reference clip Videos like this made it feel sci-fi. Functionally, this was the direction.
One operating layer, multiple workflows.
Dedicated machine, bounded tools, worker agents, and enough memory to carry work across sessions.
What it enabled
Persistent context, background work, and the operating layer behind the CRM and research systems in the rest of the talk.
Case study one
In a high-volume printing business, the bottleneck was context.
The old workflow was Zendesk plus a lot of manual reconstruction. Tickets were organized, but the actual intelligence still lived outside the workflow.
Volume made the small gaps hurt.
Once the order count is real, every manual step in support and escalation starts compounding.
Why the workflow broke
- 100,000+ customers At this volume, small workflow issues stop being small. They compound fast.
- Manual AI layer Support still depended on copying incoming emails into a separate custom GPT built around policy rules.
- Human bottleneck Routing to escalations or designers was still manual, so the drag was reconstruction, not effort.
Workflow leverage
Design principle
Design principle
Workflow mattered more than model choice.
The biggest gain came from tightening the workflow: route first, assemble context, draft in the thread, and make the system measurable.
Route first
Sort the queue, surface urgency, and send tickets where they belong before asking the model to write anything.
Draft second
Generation is most useful after context is assembled, not as a substitute for it.
VA triage, escalation context, AI reply.
Not a prettier inbox. A workflow that could route, surface context, and draft inside the thread.
Case study two
89search: a better workflow for legislative research.
The default TLO search was fine for lookup, but weak for real research. 89search made it easier to search by topic, filter by member status, and stay close to the source material.
Topic filters
Member context
Visible sources
Structured search beats blind keyword hunting.
Topic filters, member context, and visible sources made research faster than basic TLO search.
Lessons learned
01
02
03
04
What mattered most.
The useful version of this story is not model magic. It is system design under real constraints.
Narrow scope beats broad ambition
A narrow system with a real job usually beats a general system that claims it can do everything.
Context quality beats prompt cleverness
A lot of failures come from bad retrieval, weak constraints, or the wrong tool boundary, not from weak prompting style.
Humans stay at the risk boundary
Approval, escalation, and ambiguous edge cases should stay legible and easy to interrupt.
The win is lower coordination tax
The highest-value systems do not just write text faster. They reduce handoff loss, tab hopping, and repeated explanation.
Useful systems get designed, not merely prompted.
The win came from boundaries, context, review, and feedback loops, not from pretending the model was magic.