Enforcing PHI Handling Policy at LangGraph Subgraph Handoffs in Utilization Management
Treat every subgraph handoff, model call and tool call as a point where a minimum necessary decision is made for the receiving agent's purpose. LangGraph state design can limit what a subgraph is given, but allow, redact, block or escalate decisions must be enforced in the execution path at model and tool-call boundaries, and each decision must produce evidence of what the agent was permitted to receive and what actually crossed.
Placing enforcement in the execution path
- 1
Allow
The payload matches the agent's defined fields and the destination is approved. Record the decision and what was sent.
- 2
Redact
Remove or mask fields the receiving agent does not need, then forward the reduced payload. Record what was removed.
- 3
Block
Stop calls that carry out-of-scope PHI or target an unapproved destination. Return a defined error to the graph so it can handle the failure.
- 4
Escalate for human approval
Pause the call until a reviewer approves or denies it. Record the approver and the outcome.
The evidence a handoff should produce
- Agent identity and the purpose it was assigned
- Destination model or tool, and whether it is covered by a business associate arrangement
- Action taken: allow, redact, block or escalate
- Policy version and the minimum necessary basis for the agent's field set
- Timestamp and a description of what actually crossed, including fields removed by redaction
- For escalations, the approver and the outcome
Where PHI moves in a multi-agent utilization management design
A prior authorization or clinical review workflow built as a multi-agent graph typically has an intake agent, an eligibility or benefits agent, a clinical criteria agent and a determination-support agent. Each one needs a different slice of the member record. The risk is that the graph carries the whole record to all of them.
PHI travels through five places. First, shared state keys, where every node or subgraph that can read a key can read the PHI in it. Second, subgraph inputs and outputs, including any transformation applied when a parent graph invokes a child. Third, handoff payloads that one agent composes for the next, often as free text. Fourth, checkpointers and other state stores, which persist state so a run can resume or be inspected and therefore hold PHI at rest. Fifth, model and tool calls, where a prompt or a tool argument sends PHI to a model provider, an API or a retrieval system.
How much PHI crosses each of these depends on how the graph is designed. A subgraph that shares keys with its parent sees whatever those keys hold. A subgraph with its own input and output schema sees only what the mapping gives it. Neither design controls what an agent writes into a model prompt or a tool argument. That is why state design is necessary but not sufficient.
Which HIPAA requirements apply at those points
The Privacy Rule requires a covered entity to make reasonable efforts to limit uses, disclosures and requests for PHI to the minimum necessary to accomplish the intended purpose. It applies to health plans, clearinghouses and certain providers, and it requires appropriate safeguards for PHI. For a payer, that standard has to be applied to automated agents as well as to human roles. In practice, each receiving agent needs a documented purpose, a defined set of PHI fields that purpose requires, and a record of the basis for that determination.
The Security Rule technical safeguards at 45 CFR 164.312 apply to covered entities and business associates acting in accordance with 164.306. Persisted checkpoints and state stores are storage that may hold ePHI, so they belong in the same safeguard analysis as any other store. Audit controls under 164.312(b) are the reason handoff evidence matters. The security management process standard adds a further expectation: policies and procedures to prevent, detect, contain and correct security violations. A review of logs after the fact addresses detection at best. Prevention requires a decision before the data moves.
Business associate arrangements also reach into the agent stack. A covered entity may disclose PHI to a business associate when it has satisfactory assurances, in a contract or other written arrangement, that the information will be safeguarded. Any model provider, tool vendor or hosting service that can receive PHI from an agent needs to be covered, including subcontractors. A handoff that sends PHI to an uncovered endpoint is a contract problem as well as a technical one. These requirements are current as of November 2026.
What the framework provides and what must be added
Framework features help shape the problem. State schemas define what a graph holds. Subgraph scoping, where the design uses separate schemas, narrows what a child receives. Checkpointing makes runs resumable. These are design-time structures, and they are worth using deliberately: define a minimal schema per subgraph, keep PHI out of keys that do not need it, and decide whether a given checkpointer should store PHI at all or hold references to it.
What a regulated environment must add is a runtime decision on individual interactions and tool calls, applied consistently regardless of how a developer wrote a node. A schema cannot stop an agent from placing PHI into a tool argument, and a checkpointer cannot tell whether a prompt went to a covered vendor. Those checks need to run in the path of each call.
Trussed is one way to supply that layer. It enforces policy in the execution path across models, agents, tools and MCP servers, with allow, redact, block, rate limit and escalate-for-human-approval actions on individual interactions and tool calls, and it generates audit-ready evidence from that enforcement, mapped to HIPAA among other frameworks. It can be deployed self-managed in the customer's cloud with configurable retention and data residency. It does not replace state design, and it does not provide end-to-end tracing of an agentic workflow.
Implementation decisions and tradeoffs
Start by defining, for each receiving agent, the PHI fields its purpose requires, such as eligibility verification or clinical criteria review. Write down the basis. This document is the input to both schema design and runtime policy, and it is the artifact a privacy officer will want to review.
Redaction improves minimum necessary alignment but can degrade agent output if a needed field is removed, so test both. Escalation adds a human step and latency, so reserve it for cases where policy genuinely cannot decide. Blocking is the safest default for unapproved destinations but needs graceful failure handling in the graph. Persisting full state in a checkpointer is convenient for debugging and expands the PHI footprint, so decide it deliberately and apply retention limits.
- Test boundary enforcement with synthetic PHI before production, and repeat when the workflow, agents or vendors change
- Confirm business associate coverage for every model, tool and hosting provider that can receive PHI, including subcontractors
- Align agent policy with the security management process for prevention, detection, containment and correction
- Apply minimum necessary policy to automated agent access, not only to human roles
- Ask of any design: is enforcement in the execution path, or only in prompts and after-the-fact log review?
One handoff, four control points
Subgraph input
Reduced payload scoped to the receiving agent's purpose
Policy decision
Allow, redact, block or escalate for human approval
Model or tool call
PHI leaves the boundary only if policy permits it
Evidence record
Permitted versus actual, tied to a policy version
Review your handoff boundaries
Trussed enforces policy in the execution path at model and tool-call boundaries and produces audit-ready evidence for regulated enterprises, including healthcare organizations running in production. Walk through your utilization management design with an expert.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.