One governed model
People and agents need a reliable shared understanding of customers, work, permissions and state.
05 / Built for people and agents
The next user of your software might not be a person.
Most software assumes a human will log in, find a screen, click through it and move the answer somewhere else. Agents change that.
Agents need governed tools and permissions, not another folder of PDFs.
Tools with boundaries
An agent should know what it can read, what it can change and when a human needs to decide.
Autonomy gradually
Start with useful assistance. Increase autonomy only where the evidence says it is safe and valuable.
HERE'S WHAT BETTER COULD LOOK LIKE
Give agents a bounded job, not the keys to the building.
Define what an agent can understand, retrieve and do, then make the handoff to a human explicit where judgement or approval is required.
“Where is my application?”
01Identify customer
02Read permitted status
03Check outstanding tasks
04Prepare accurate response
05Act if permitted
06Ask a human when judgement is needed
CAN SEEstatus · tasks · permitted records
CAN DOdraft · update · create · request
HUMANjudgement · exceptions · approval
WHAT THIS LOOKS LIKE
Design for action, not just answers.
Understand
Give the agent a reliable model of customers, work, state and context.
Retrieve
Let it access the information needed for the job without giving it the keys to everything.
Act
Expose explicit tools such as create task, update record, prepare handover or request approval.
Escalate
Make the point where human judgement is required part of the architecture, not an afterthought.
What is one recurring job you would genuinely let an agent move forward if the boundaries were clear and a human remained in control?
Connected software creates room for useful agents.
Yarta already works across connected email, calendar, tasks and notes. That same principle matters inside organisations: agents become more useful when the underlying information and actions are deliberately connected.
See the work →4 specific ways in.
Each chapter starts with a real situation rather than another abstract software claim.
WEEK 17
The next user of your software might not be a person
Most software assumes a human will click through it. Agents change that.
See the story →WEEK 18Agents need tools, not another folder of PDFs
Knowledge helps. Permission to act is what changes the work.
See the story →WEEK 19AI is easy. Permission to act is harder.
The interesting question is not what the model can do. It is what you are prepared to let it do.
See the story →WEEK 20Where would you actually let an agent do something?
One recurring job is enough to find out.
See the story →Bring us one recurring job.
Pick something boring enough to be safe and useful enough to matter. We will map what an agent could see, do and escalate.
