How to Scope PHI Between Agents in Amazon Bedrock Multi-Agent Collaboration
Scope PHI by treating each agent as a role with its own permitted data classes, inputs, outputs and tools, then enforce those limits at runtime at four points: routing, context passing, tool calls and returned results. IAM, agent configuration and instructions set the design-time boundary. They do not evaluate the content of an individual interaction, so allow, redact and block decisions need to be made in the execution path, and each decision should be recorded as evidence.
Where to apply allow, redact, block and escalate
- 1
1. Routing
Confirm the target collaborator is permitted to handle the task type. Block delegation to a collaborator outside the request's scope.
- 2
2. Context passing
Compare the outgoing context with the collaborator's permitted inputs. Redact data classes it does not need. Block if the task cannot proceed without them.
- 3
3. Tool and action group calls
Evaluate tool inputs against the agent's permitted tools and data classes. Escalate sensitive calls for human approval before execution.
- 4
4. Tool outputs
Check what the tool returned before the agent uses it. Redact fields outside the agent's permitted use.
- 5
5. Returned results
Filter the collaborator's response against its permitted outputs before it enters the supervisor's shared context.
How PHI moves in a supervisor and collaborator design
In Amazon Bedrock multi-agent collaboration, a supervisor agent routes tasks to collaborator agents and passes context, instructions and results between them. In a healthcare deployment, that flow is also a data flow. A patient request that reaches the supervisor may contain identifiers, clinical details or coverage information, and the supervisor can forward some or all of it when it delegates.
The risk runs in both directions. PHI can move from the supervisor to a collaborator that has no need for it. It can also come back: a collaborator's response, or the output of a tool it called, can carry PHI into the supervisor's shared context. From there it can reach other collaborators, other tool inputs and the final response to the user.
This gives four interaction points to design around: routing, context passing, action group or tool calls, and returned results. Each is a place where PHI can enter, persist or be forwarded, so each needs an explicit scoping decision. Returned results deserve separate treatment from delegation inputs. Teams often design the outbound path carefully and leave the return path open.
Design-time controls and runtime controls are different things
Design-time controls include IAM roles, agent configuration and the instructions given to each agent. They determine what an agent is able to reach and how it is told to behave. They are necessary, and they should follow least privilege. An agent whose role cannot reach a data store cannot read from it.
They have limits in a multi-agent setting. IAM scopes access to resources, not the content of a particular message. Once a supervisor holds PHI in its context, a role policy on a collaborator does not stop that PHI from being included in a delegation. Instructions such as 'do not share patient identifiers' are probabilistic. They depend on model behavior and can fail under unusual inputs or adversarial prompts.
Runtime controls evaluate individual interactions and tool calls as they happen. They can inspect what is about to be sent to a collaborator, what a tool is about to receive, and what a collaborator is about to return, then allow, redact or block. Agent permissions are a recognized exfiltration risk elsewhere: research on another vendor's agent platform has described default permission scoping being abused by a compromised agent. The lesson is portable. Do not assume an agent's permissions are as narrow as its task.
Define a per-agent PHI scoping model
Start with an inventory of what each agent role needs to do its job. Then write a matrix with one row per agent. Each row should state the PHI data classes the agent may receive, the classes it may use, the classes it may return, and the tools it may call. The regulatory anchor is the minimum necessary standard, referenced at 45 CFR 164.502(b). The practical consequence is that a collaborator should receive what its task requires and nothing more.
The matrix below is illustrative. Adjust the classes to your own data model.
- Supervisor: Receives user input. Holds the minimum PHI needed to route. Returns only the final response, with PHI limited to what the user is entitled to see.
- Scheduling collaborator: May receive name and appointment details. Must not receive diagnosis, clinical notes or billing data. Tool access limited to the scheduling system.
- Coverage collaborator: May receive member identifier and plan details. Must not receive clinical free text. Returns eligibility status, not full member records.
- Clinical summarization collaborator: May receive clinical content for a single encounter. Returns a summary with direct identifiers removed unless the supervisor's scope permits them.
- Any agent with write-capable tools: Calls that change records or send data outside the organization require human approval.
Evidence to retain for HIPAA
A scoping model is only useful in an audit if you can show it was applied. HIPAA's technical safeguards at 45 CFR 164.312 include audit controls at paragraph (b). Separately, the incidental-use provision in Part 164 is conditioned on compliance with 164.502(b), 164.514(d) and 164.530(c). As of November 2026, all of these sections remain in the current eCFR, and no amendment or repeal has been reported in the material reviewed.
For agent systems, a reasonable approach is to retain a record per runtime decision. Each record should capture the agent identity, the interaction point, the data class involved, the action taken (allow, redact, block or escalate), the policy version in force and the time. For escalations, add who approved and when. Together with the per-agent matrix itself, this lets you show what each agent was permitted to receive, use and return, and what actually happened.
If you use limited data sets, check downstream arrangements as well. The limited data set provisions in 45 CFR 164.514 require that any agents to whom the recipient provides the data agree to the same restrictions and conditions. Contract terms and agent configuration should be consistent on this point.
Separate configuration evidence from interaction evidence. IAM policies and agent definitions show intended scope. Per-interaction records show enforced scope. Reviewers will usually want both.
Tradeoffs and where Trussed fits
Runtime enforcement adds latency and policy maintenance. Redaction can degrade a collaborator's output if it removes context the task needed, so test each policy against real task flows. Start with the highest-risk paths: tools that write or transmit data, and collaborator responses that feed the supervisor's shared context. Tighten the matrix as you learn what each agent actually uses.
Trussed enforces policy in the execution path across models, agents, tools and MCP servers, at the level of individual interactions and tool calls. The actions are allow, redact, block, rate limit and escalate for human approval. Enforcement is at runtime, not reviewed after the fact, and it produces audit-ready evidence mapped to HIPAA. Trussed runs in production with healthcare and pharma organizations and can be deployed self-managed in the customer cloud, with configurable retention and data residency.
A sensible next step is to draft the per-agent matrix for one workflow, mark which boundaries currently depend only on IAM or instructions, and decide which of the five decision points above need runtime enforcement first.
Where PHI scoping decisions apply
Routing
Decide which collaborator receives the task and which PHI classes it may see.
Context passing
Redact or block PHI the collaborator does not need before delegation.
Tool calls
Check inputs and outputs of action group and tool calls against the agent's scope.
Returned results
Filter collaborator responses before they enter the supervisor's shared context.
Review your agent PHI boundaries
Walk through a supervisor and collaborator workflow with us and see how runtime enforcement and audit-ready evidence apply to each interaction point.
Explore Agent and Tool-Call EnforcementHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.