Requiring Human Confirmation Before Google ADK Agents Place Clinical Orders
Classify every tool that creates or modifies a medication, lab, imaging or referral order as consequential, then require a named clinician to approve each call before it executes. Confirmation in ADK tools and callbacks is a valid first layer. For clinical orders, add a policy enforced at runtime in the execution path, so approval does not depend on each agent's own code, and retain a record of who approved what, and when.
Evidence to capture for HIPAA audit controls
- Approver identity and the authentication context
- The order parameters as presented to the approver
- The decision: approved, denied or expired
- Timestamps for the request, the decision and the execution
- The outcome returned by the downstream system
- The agent and tool that proposed the call
Identify which tool calls are clinical orders
Start from the tool inventory, not from the agent's intent. Any tool that can create, change or cancel a medication order, a laboratory order, an imaging order or a referral is a clinical order tool, whether it wraps a FHIR resource write, an EHR vendor API or an internal service. Read-only tools that retrieve results or history are a different class and usually do not need per-call approval.
Classify by effect rather than by name. A tool called "update_encounter" may place an order if one of its parameters does. Review each tool's write paths and arguments, and record the classification alongside the tool definition so it can be reviewed. Treat any tool that has not been classified, and that can write to a clinical system, as consequential until someone decides otherwise.
Assign ownership. Someone in clinical informatics or security should own the list of consequential tools, and changes to that list should follow change control. A classification that lives only in a developer's head will drift as agents and tools are added.
Confirmation at the ADK tool and callback level
Within an ADK agent, confirmation can be implemented in two general places: inside the tool function itself, or in a callback that runs before the tool executes and can allow, modify or stop the call. A tool can also be designed to pause and wait for an external decision before returning. Each is a legitimate way to put a human in the loop, and for a prototype or a single well-owned agent it may be enough.
The limits are structural. The control lives in application code, so it is only as consistent as the teams writing it. A new agent that reuses a tool without the callback, a tool whose confirmation flag is misconfigured, or a prompt-injected agent that reaches an order tool by another path can all skip the prompt. Each agent also ends up with its own approval screen, its own identity handling and its own records, which makes a uniform audit story hard to assemble.
ADK-level confirmation compared with runtime policy enforcement
The two approaches are complementary. The question is where the guarantee comes from. For a clinical order, the stronger guarantee is a policy that sits in the execution path and applies to every call, whichever agent made it.
- Where the control lives: ADK-level: in each agent's tool or callback code. Runtime policy: in the execution path, outside the agent's own code.
- Consistency across agents: ADK-level: depends on each team's implementation. Runtime policy: one policy applies to every matching tool call.
- Exposure to misconfiguration or injection: ADK-level: a missed callback or altered flow can bypass the prompt. Runtime policy: the call is held at the point of execution regardless of how the agent reached it.
- Approval evidence: ADK-level: format and retention vary per agent. Runtime policy: one record format, produced by the enforcement itself.
Design the approval step
Whichever layer you use, the approval step determines whether the control is meaningful. Approval fatigue is a real risk: if clinicians are asked to confirm many low-value actions, they will approve without reading. Limit prompts to consequential calls and show only what the decision needs.
Bind the approval to the exact call. If arguments can change between approval and execution, whether through agent retries or manipulated parameters, the clinician has approved something other than what runs. Execute only the arguments that were displayed, or invalidate the approval if any argument changes.
- Approver identity: authenticate a named clinician through your existing identity provider, and check that the person is authorized to place that order type. A shared service account is not an approver.
- Order context shown: patient identifier, order type, drug and dose or test and indication as applicable, the requesting agent, and the arguments exactly as they will be submitted.
- Denial behavior: a denial stops the call, returns a clear result to the agent so it does not retry the same order, and is recorded.
- Timeout behavior: set an expiry on pending requests and fail closed. An unanswered request is not an approval.
- Single use: one approval covers one call. Do not let an approval carry over to later calls in the same session.
Enforcing escalation in the execution path with Trussed
Trussed enforces policy on individual interactions and tool calls in the execution path. For this use case, a policy can match calls to clinical order tools and escalate them for human approval before they execute. Other policies can allow, redact or block, so read-only calls proceed without interruption. Because the policy is enforced at runtime, not reviewed after the fact, it applies whether or not a given agent includes its own confirmation logic.
The enforcement produces audit-ready evidence, mapped to HIPAA among other frameworks, so the approval record comes from the control itself rather than from each agent's logging. Trussed can be deployed self-managed in your cloud, with configurable retention and data residency, and it runs in production with healthcare and pharma organizations.
This addresses individual tool calls. It is not a substitute for classifying your tools or designing the approval experience, and it does not make claims about clinical outcomes. Keep ADK-level confirmation as a first layer, and use runtime enforcement as the layer that does not depend on any one agent being built correctly.
Approval before execution
Agent proposes
The agent selects an order tool and prepares arguments.
Policy evaluates
The tool call is classified as consequential and held.
Clinician decides
A named approver sees the order details and approves or denies.
Call executes or fails closed
Only an approved call proceeds. Denial and timeout do not.
Evidence retained
Approver, order, decision, time and outcome are recorded.
See how approval is enforced and evidenced
Review how escalate-for-human-approval policy applies to individual tool calls and how the resulting evidence supports HIPAA audit controls.
Explore Agent and Tool-Call EnforcementHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.