Agent Delegation Needs a Work Contract

Multiple AI agents can break down a complex process, gather information, and combine results. Google Cloud currently describes “contract-first decomposition”: work is split so that delegated results remain verifiable. At the same time, Anthropic has made Computer Use, Browser Use, the Skills API, and the Files API generally available for production agents. This increasingly turns a demonstration into an operational process.

Delegation is not an open-ended prompt

An orchestrating agent should not simply say, “Handle this case.” Each subtask needs a clear frame: objective, permitted data, allowed tools, expected output, cost limit, and acceptance criterion. Only then can a company determine whether a sub-agent has delivered or merely produced a plausible-sounding response.

Google highlights two further reasons. First, a smaller model may be more economical for a simple subtask than a powerful reasoning model. Second, sensitive data should be passed on only with the minimum necessary permissions. Both become manageable when the subtask states its boundaries explicitly.

New tools increase responsibility

Anthropic’s announcement covers components that let agents operate software, use versioned capabilities, and process files. That expands the value of agents, especially in processes with legacy interfaces and documents. It also expands the number of decisions an agent can prepare or trigger.

A work contract therefore separates four levels:

Outcome: What decision or artefact is expected?
Authority: Which data, applications, and actions are permitted for this task?
Budget: Which model, token, or time limit applies to the subtask?
Verification: Which rule, metric, or human role accepts the result?

What DACH organisations should establish now

Start with a process whose outcome is already reviewed, such as preparing a credit application, analysing an incident, or summarising a contract. For two or three subtasks, define the input, tool scope, output schema, and acceptance criterion. Also log which agent received which data, which model was used, and who approved the result.

This is not another bureaucracy layer around a prompt. It is a controllable delivery chain for AI work. The key question is no longer how many agents run in parallel, but whether every delegated step has a verifiable assignment and an accountable recipient.