Hand big agent work to a licensed Claude worker
When a job is too big for an agent's own model turn, the agent hands it to a person with a Claude licence. That person works it in Claude Code, the Agent SDK or a managed agent, through a governed connection that only reaches what the agent itself may reach.
An agent normally does its work in short model turns, each limited in how much it can read. Some work is bigger than that: a year of supplier invoices, every open quality notification, a full ledger review. A tool of the kind 'Hand to a licensed worker' lets the agent pass that work to a person who holds a Claude licence. The run waits until the result comes back.
- The agent calls its 'Hand to a licensed worker' tool. The tool's grant, guardrails and any approval are checked first, exactly as for any other tool.
- The run pauses with the status Waiting for input, and a task appears under Agentic AI > Operate > Delegated Tasks with a DLG number.
- A person with the Heavy Worker role opens the task and chooses Claim. The result panel shows a worker token and the connection to paste into Claude Code, the Agent SDK or a managed agent. Copy it now: it is shown once.
- Their Claude session reads the task with get_delegation, does the work with the agent's own tools, attaches any files, and finishes with submit_result. Submit result on the screen does the same.
- If the tool asks for a review, a reviewer accepts or rejects the result. Otherwise the run picks the result up straight away and carries on.
- The worker token belongs to the agent's identity, not to the worker. It lasts an hour, can be renewed while the task is open, and works only against the TRANXACT MCP address.
- The token stops working the moment the task is submitted, released, cancelled or expires, or when the run ends.
- Every call the worker makes is checked against the agent's grants, guardrails and approvals and recorded as a tool call, with the worker's name beside the agent's in the audit trail.
- The card vault is never reachable, and identifier fields such as national IDs and bank account numbers are withheld from every result.
Names, e-mail addresses and similar details in the records the agent may read are visible to the worker's Claude model. The masking that in-app agents get does not apply outside TRANXACT. Keep the agent's grants narrow. To have a person approve every hand-off, mark the delegate tool as touching personal data.