Vinicius Apolinario
de Brazil
Inspired when he saw a 1 gigabyte hard drive
Jumpers on hardware to change settings
Virtual servers
VMware, virtual box
Cloud
Containers > kubernetes
AI > MCP > agents > claws >
All new technologies are disruptive
The job itself did no disappear
The abstraction moves one layer up
AI is moving faster and faster
The task that the AI agent is going to do already exists
AI agents have goals
Users are going to put a lot of stuff in production that operators do not necessarily trust
What are the opportunities ?
What are the responsibilities ?
Reduce repetitive work
Reduce the stress of problem-solving
Improve the processes
This all needs adult supervision!
Agents can act with incomplete context
They go off in an odd direction…
Your AI agent can move quickly in the wrong direction
The formula:
Context - what really matters
Boundaries - what is allowed and the limits
Feedback - whether the agent did something right or wrong
Accountability - if an AI is doing a process, someone is held responsible
Do not automate a process unless you understand it!
What problem are we solving?
Who experiences the pain?
What data matters?
What does “good” look like?
What happens if it is wrong?
Should this be automated at all?
Strategic thinking should probably have a human involved
Lawyers do not know the answers to the questions, which are nuanced
Boundaries
Identity
Permissions
Tool access
Data access - does the agent have its own access or that of a human?
Approval flows
Network boundaries
Secrets - who is asking for the data?
Cost limits (this is an specific AI problem for subscriptions)
suggest (low risk) > draft (human reviews) > execute with approval (human authorizes) > execute within limits (guardrails + monitoring) > autonomous (strong controls + rollback)
/whatif explores without executing
Feedback
traditional signals: latency, errors, dependencies, cost, availability
Agent signals: goal, prompt/context, retrieved data, tool calls, evaluation result
Human signalrs
User feedback, overrides, approvals, excavations, business outcomes
If you cannot observe, you cannot improve.
Accountability
Who approved the workflow?
Who owns the data?
Who owns the tools?
Who responds when it fails?
Who decides the autonomy level?
Who explains the outcome?
How is a system behaving on behalf of people?
A person would have accountability. An agent needs this ,too.
Developed debate:
The impact of coding agents on developers
Junior developers do not have the experience to understand if generated code is the best
If the code executes what I want, isn’t that okay
The value of the developer is in understanding the value of the code, not in writing it.
Infrastructure:
AI is just another workload.
Access. Logs. Cost. Events. Typical and known issues.
What is the blast radius when something goes wrong?
Backups? Restoration?
Data:
Context is key.
Accessing the data is not the same as understanding the business.
What is the business goal?
Business Leader:
Pressure comes from shareholders > leaders > managers > workers
“Do AI”
Implement AI to make it better!
AI should not necessarily change the process.
When are decisions slow? Where can we improve the process?
What work is repetitive? What are the low-hanging fruits? What needs to be studied?
Homework
Build one small agent
Connect it to one real workflow
Add one meaningful boundary
Measure one outcome
Prioduction
1. Pick one painful
2. Indentify safe agent assistance
3. …