If your software cannot be used by your agents, it is not ready for the next operating model.
Most companies still build software for one user: a human sitting in front of a screen.
That made sense for the last era of work. People logged in, clicked through menus, copied data between systems, chased approvals, checked dashboards, and updated records manually.
That is not where work is going.
If we are serious about digital coworkers, agentic workflows, and AI moving from chat to execution, we need to change how we design applications. Every application should answer a more demanding question than “Can a person use this?”
Can an agent operate it safely?
That is the standard we are applying internally now. We no longer think of applications as tools that only humans touch. Everything we build needs to be usable by the agents that support our work.
I put a simple readiness checklist in the gist Can This App Be Agentified?. It helps us decide whether an application can be operated by agents or needs remediation first.
A polished UI is not enough
A lot of teams confuse usability with operability.
A clean interface helps humans. It does not automatically help agents. An agent needs a reliable way to understand the system, take action, verify the result, and escalate when risk crosses a threshold.
Every app therefore needs more than screens. It needs a control surface: an API, service layer, workflow endpoint, structured database interface, event queue, or set of safe actions exposed to agents.
Without one, the agent becomes a fragile screen-clicking assistant. It has to guess where buttons are, read changing layouts, and rely on UI state instead of system state. It breaks when someone changes a label or moves a modal.
That is automating around bad architecture.
The six checks we use
The checklist breaks agent readiness into six signals.
1. Resource model
Can the app clearly describe the things it manages?
Those things might be a customer, invoice, ticket, booking, task, lead, approval request, or training enrolment.
If the app has no clean resource model, the agent has no stable object to work with. It may see text on a page, but it does not know what the object is, what state it is in, or which actions are valid.
Humans can infer messy context. Agents need structure.
2. API or control surface
Can the agent take action without relying only on the UI?
Many applications fail here. They have a decent interface for people but no practical way for an agent to create, update, retrieve, or trigger anything safely.
If every workflow depends on clicking through screens, the app is not agent-ready yet. It may still be useful, but it needs remediation before agents should operate it at scale.
3. Permissions
Can we define exactly what the agent is allowed to do?
Agents should not operate with vague access. A sales agent may update CRM notes but should not change contract terms. A finance agent may prepare an invoice but should not approve payment. A training operations agent may check enrolment status but should not override eligibility rules.
Agent-ready applications need permission boundaries that match real business authority. Not every action should be available to every agent.
4. UI blockers
Are there steps that still require a human to stare at a screen?
Some blockers are obvious:
captchas
hidden modals
drag-and-drop-only actions
unstructured forms
unpredictable popups
workflows with no system readback
Others are more subtle. An app may require users to visually compare two records before confirming a change. That can work for a human operator, but it is poor for an agent unless the comparison data is exposed clearly.
The UI does not always need to disappear. The underlying workflow needs to be exposed in a form an agent can use safely.
5. Trust gate
What happens when the action is risky?
Agent-ready does not mean fully autonomous for everything. Some actions should still require approval. Approval should be designed into the workflow instead of handled through random Slack messages and manual screenshots.
A good trust gate tells the agent when to proceed, when to ask, who can approve, what evidence to show, and what to log after the decision.
That is how you keep humans in the loop without keeping humans in every loop.
6. Audit and readback
Can the agent verify what happened after it acts?
This is non-negotiable. If an agent updates a customer record, it needs to read back the record and confirm the change. If it triggers a workflow, it needs to know whether the workflow started, failed, or completed. If it escalates for approval, it needs to know who approved, when they approved, and what was approved.
Without readback, you do not have reliable execution. You have a request sent into the dark.
Design operating surfaces, not just features
Most software teams still think in terms of features: what can the user do?
Agentic companies need to design operating surfaces: what can a human, agent, or workflow safely coordinate through this system?
That shift changes product design. Schemas, permissions, logs, event states, rollback paths, and exception queues become first-class concerns. Screens that look complete but still require a person to copy, paste, chase, check, and update are no longer enough.
That is the work we want to remove, not because humans do not matter, but because humans should not be trapped operating software that should be operating with them.
Build for digital coworkers from day one
The next generation of applications will be judged by how well they can be orchestrated.
Can agents understand the system? Can they act through safe interfaces? Can they escalate risky decisions? Can they prove what happened? Can they work alongside humans without turning every workflow into another manual checkpoint?
If an app cannot be used by your agents, it is not ready for the next operating model.
Use the agent-readiness checklist on one application this week. If it fails one of the six checks, you have found the next remediation task.
The Nexius guide to connected AI applications and systems shows how agents can work with ERP, CRM and operational systems without weakening the authoritative system of record.
Top comments (0)