CLI
made software commandable.
The derivation
Where HAI comes from
CLI, TUI, GUI, HAI: every new tool class needs a new interface.
Agents can research, plan, build, verify, and coordinate. They do not only wait for input. They act. That changes the interface problem from input to direction.
Interface Shift
Every new tool class needs a new interface.
Commanding software, navigating systems, controlling computers, and directing agents are not the same act. HAI is the name for the fourth act.
TUI
made systems navigable in text.
GUI
made computers visually controllable.
HAI
is for directing systems that can act.
When systems can act
The interface has to hold intent, scope, risk, feedback, and control.
Raw agents amplify unclear intent into unclear work. Human-Agent Interface exists to make the work legible before the agent starts moving.
Messy input
An idea, a blocked project, a vague decision, ten open threads.
Raw delegation
The agent runs, expands scope, creates artifacts, and the human loses the thread.
Agentic headache
More output appears, but the next human action is still not clear.
HAI is derived from the moment software stops being passive.
It is not a single chatbot, dashboard, or auto-agent. It is the translation and control layer between human chaos and systems that can act.
Mechanism
The product kernel is a controlled work packet.
HAI takes a human state and turns it into a bounded assignment: role, scope, risk, artifact obligation, and next step. The loop is stable; Human Fit explains what adapts to the person using it.
Capture the human problem state without pretending it is already structured.
Name what this actually is: idea, blockage, project state, decision, or risk.
Separate agent-capable work from what must remain human-owned.
Choose the right operating mode instead of handing everything to one generic assistant.
Specify what must exist at the end so progress can be inspected.
Produce a concrete packet the human can approve, run, reject, or revise.
Agents are roles, not magic.
A useful interface routes work into a mode the human can supervise. Sometimes that means no agent at all.
Collects evidence, maps options, and reduces uncertainty before action.
Performs bounded implementation work against a clear definition of done.
Checks whether the promised outcome actually holds in reality.
Coordinates multiple agents while preserving state, ownership, and handoff clarity.
Marks work that depends on judgment, consent, taste, risk appetite, or irreversible decisions.
Control
The human does not disappear from the system.
The point is not autonomy theater. The point is controlled working ability with agents.
The human owns
- Goal and meaning
- Scope boundaries
- Risk and approvals
- Irreversible decisions
- The final sense of whether the work matters
The agent works on
- Research and synthesis
- Implementation and iteration
- Verification and critique
- Documentation and handoff
- Structured next actions
Human tool use continues
The agent is not the product. The collaboration is.
Agents may be the most powerful tools humans have ever had. But power without an interface becomes noise. Human-Agent Interface is the discipline of turning that power into clear, supervised, useful work.