AAP has three parties: the agent, the adapter and the provider.
Agent
The agent is the system that wants to call a tool. A harness runs the agent and executes its tool calls.
The agent does not need to know how approval is obtained. It makes a tool call and receives either the tool's result or an explanation of why the call did not run. AAP is transparent to the agent.
Adapter
The adapter sits between the agent and its tools. It intercepts a tool call before execution and asks the provider whether it may run.
The adapter is responsible for enforcing the answer. It must prevent the call from running until it has received approval for that call.
Provider
The provider decides whether a call may run. It records the request and returns the result to the adapter.
A decision may be immediate, for example when a policy approves the call automatically. It may also require a human review and take much longer.
Modes
Human review can take hours or longer. AAP supports synchronous and asynchronous modes to accommodate different harness capabilities.
In asynchronous mode, the harness saves the pending tool call and suspends execution. When a decision is available, it resumes execution and the adapter applies the result.
This means the agent process does not need to remain running whilst waiting for approval. The pending call can also survive a process restart, provided the harness has saved enough state to resume it.
Synchronous mode supports harnesses that wait for a tool call to return before continuing execution. The adapter holds the tool call open, submits an approval request and polls the provider until a decision is available.
Both modes use the same requests and decisions. They differ in how execution waits for the result.