Skip to main content

Governing multi-agent orchestration: delegation, memory and control

Several agents can investigate a problem, gather information and carry a business process through multiple steps. The coordination becomes useful when each participant has a clear job and a controlled way to act. Security has to survive every delegation, memory read, approval and retry that happens between the first request and the final outcome.

In this article

Give the team a business contract

Begin with the outcome the team is expected to produce. Define the entry conditions, completion criteria, accountable human owner and decisions reserved for people. Then describe the contribution of each agent. A research role, a reviewer and an execution role should have different inputs, permissions and outputs.

Document the handoffs as carefully as the individual tasks. Who receives an incomplete result? Which source is authoritative when agents disagree? How long can the coordinator wait? Which failures require human intervention? These decisions turn a collection of conversations into an operation that can be monitored and recovered.

Keep the architecture proportionate. Add another agent when a distinct responsibility or permission boundary justifies it. More agents introduce more state, communication and failure combinations to manage. A workflow with a few well-defined roles can be easier to control than an open-ended team that repeatedly decides what to do next.

Delegation must preserve the authority boundary

A coordinator needs permission to allocate work. It should not acquire the ability to grant whatever access would make that work convenient. Define the tools, actions, resources, purposes and operating conditions that may be delegated. A recipient’s effective permissions must also respect its own identity and the current policy.

For example, an investigation agent might be allowed to retrieve specific customer records and ask a drafting agent to prepare a reply. That delegation does not permit the drafting agent to search unrelated accounts, send the response or issue a credit. Each of those actions has its own authority requirements.

Make delegation explicit and reviewable. Record the parent, recipient, task scope, expiry and relevant budget. A revoked parent grant should have defined consequences for outstanding child work. Limit further delegation where it is unnecessary, and ensure cancellation reaches work that is queued as well as work already running.

OWASP’s Excessive Agency guidance identifies excessive functionality, permissions and autonomy as contributors to damaging agent actions. The practical design response is to give each role the smallest useful operating remit and enforce it outside the model’s instructions.

Treat a handoff as a request that needs validation

Passing a message to another agent is a security boundary. The receiver needs a trustworthy task identifier, sender identity, customer and workspace context, purpose and permitted scope. It should reject missing or mismatched context rather than trying to infer the correct customer from the text.

Separate the task envelope from the material being analysed. A web page, email, tool response or peer-agent summary can supply evidence, but it should not rewrite the recipient’s permissions. Preserve source references so the next participant can distinguish an observation from a hypothesis or recommendation.

Validate structured outputs before using them. A field labelled customer identifier must match the authorised case, and a proposed destination must be an approved endpoint. Free-form confidence statements from another agent are not substitutes for these checks.

Shared memory needs ownership and purpose

Giving every agent the same conversation history is convenient, but it can disclose information that some participants do not need. Design separate areas for case state, durable knowledge, operational evidence and temporary working material. Apply access and retention rules to each.

A shared case summary should identify its source, author, version and review status. Store uncertain claims as uncertain. When an agent corrects a fact, preserve enough history to understand what earlier decisions relied on. Avoid silently converting a speculative conclusion into durable organisational knowledge.

Retrieve memory within the current identity, tenant, workspace and task boundaries. Recheck access when work resumes because permissions may have changed during a pause. Keep secrets out of shared messages and logs, and decide which customer information may be sent to each model provider.

LangGraph’s persistence documentation distinguishes thread-level checkpoints from stores used across threads. Those mechanisms help organise state. The application must still implement its own authorised access boundaries and retention decisions.

Separate orchestration progress from execution permission

An orchestration framework decides which step runs next. The point that executes a consequential action needs to verify whether that specific action is currently permitted. This remains necessary when a previous step has marked the task approved or another agent claims that a human agreed.

Bind approval to a concrete request: the target, action and relevant parameters. If the amount, recipient or data scope changes, the system needs to evaluate the changed request. A generic approval flag copied between agents can lose that connection.

Human review also needs a usable interface. Show the proposed action, evidence, expected effect and unresolved issues. Identify who is authorised to decide and what happens if they reject it or do not respond. Keep the approval record separate from the model’s interpretation of the reviewer’s message.

LangGraph interrupts provide a way to pause for external input and resume work. Its documentation also explains how an external approval system can remain authoritative. A resumed graph should observe the governed job’s status, rather than treating the act of resuming as permission to execute.

A worked example: investigating a customer credit request

Imagine a hypothetical service business handling a disputed invoice. An intake agent opens a case, a records agent retrieves the permitted invoice and service history, and an analysis agent prepares a proposed response. A finance reviewer decides whether a credit is justified. An execution agent submits only the approved change through the configured connector.

The coordinator can request missing facts and keep the work moving. It cannot expand the records agent’s customer scope or authorise a credit itself. The drafting agent receives the relevant case summary rather than unrestricted finance access. The reviewer sees the invoice reference and the exact proposed amount.

Suppose an attachment tells the records agent to ignore its rules and send the account history to an external address. The attachment remains untrusted content. The intended design denies that destination independently, even if the agent proposes the send. OWASP’s prompt-injection guidance describes the risk of instructions arriving through external material.

If finance changes the credit amount, the original approval cannot simply be reused. If submission times out, the workflow checks whether the connector completed the action before retrying. If the request is cancelled, pending child tasks and unconsumed authority must be handled consistently. These are part of completing the customer’s case correctly.

Engineer recovery and limits before a busy day

Define limits for parallel tasks, child agents, tool calls, model usage, elapsed time and financial impact. A valid action may still exceed the process budget. Decide whether reaching a limit should pause the case, request additional authorisation or end the task with a clear explanation.

Persist enough state to recover safely, including the identifiers of submitted jobs and any confirmed outcomes. Distinguish a known failure from an unknown outcome. Repeating an operation whose result is uncertain can create duplicate communications or changes.

Retries need an operation-specific strategy. Idempotent endpoints and deduplication keys can help, but the receiving system must enforce the expected behaviour. LangGraph’s interrupt documentation cautions that resumed nodes can run code again. Design and test side effects with that replay behaviour in mind.

Review the whole chain of evidence

Test a normal case, a rejected handoff, a changed approval, a revoked identity, an unauthorised memory read and an interrupted execution. Follow the task across participants. The evidence should show which policy and identity applied, who approved the action, what actually executed and what remains unresolved.

Gamut Gateway’s control model covers bounded delegation, governed handoffs, scoped memory and execution evidence for supported integrations and mediated paths. Coverage needs to be established for the deployed workflow. An action that bypasses its enforcement path does not become governed because an agent appears in the register.

Gamut’s records, assessments and reports support the broader governance lifecycle. The AI Assurance Officer and expert services can help prepare and review agreed assurance work. Accountable people retain formal decisions, while deployment-specific tests establish whether the intended controls operate effectively.

A successful agent team therefore produces two connected results: the business outcome and a trustworthy explanation of how it reached that outcome. Both are needed before the organisation can responsibly extend the team’s scope.

Sources & further reading