When organisations first plan AI agents, they usually draw their org chart and put one agent in each box: a procurement agent, a finance agent, a legal agent. It looks tidy. In practice it reproduces every hand-off problem the organisation already has, now with software that cannot walk down the corridor to ask a question.
Slice by the result, not the reporting line
On a recent agentic platform for a national authority, we took the opposite route. We listed the outcomes the authority is accountable for across the life of a contract, and gave each outcome to one specialist agent. Inspection and acceptance. Revenue and payments. Penalties and compliance. Change orders. Exit and continuity. Ten agents in total, each owning one result end to end, even when that result crosses three departments.
The benefit is accountability. When a payment is late, there is exactly one agent whose job is to know why, and one human owner who reviews its recommendation. Nobody has to reconcile three departmental bots that each saw part of the picture.
Four layers keep it governable
Outcome-based agents only work on a shared base. We design the platform in four layers. At the bottom, a governed data foundation where contracts, invoices, inspection reports and correspondence live once, with clear ownership and residency. Above it, shared services every agent uses: identity, retrieval, document understanding, logging and audit. Then the specialist agents. And at the top, human decision gates where a named person approves, rejects or sends back each recommendation.
The top layer is not a formality. In the public sector, an agent may draft a penalty notice or a termination recommendation, but a person signs it. Designing the gates on day one makes audits straightforward and makes leadership comfortable going live sooner.
Start with one journey on one live contract
The temptation is to build all ten agents in parallel. We recommend the opposite: build the foundation, then run one journey on one live contract. In our case that was the execution line from award to renewal, with inspection and payments going live first because they had the cleanest data and the most visible pain.
Running one contract end to end exposes the real integration and data issues early, while the blast radius is small. It also produces the evidence the sponsor needs to fund the next journey: fewer late payments, faster acceptance, a cleaner audit trail.
A quick test for your own design
How agents hand work to each other
Outcome-based agents still need to cooperate. When an inspection is accepted, the revenue agent needs to know so it can prepare the payment. When a penalty is applied, the contribution analyst needs to know so it can update the contract's value. We handle this with events on the shared services layer rather than with agents calling each other directly. Each agent publishes what it has decided and subscribes to what it needs.
This keeps agents independent. You can improve, retrain or replace the penalties agent without touching the payments agent, as long as the events stay the same. It also gives auditors a clean timeline: every event is logged with the agent that raised it, the evidence it used and the human who approved it.
Where external research belongs
Some outcomes need information from outside the organisation, such as market prices, supplier news or regulatory changes. We keep that research behind a wall. A dedicated agent gathers and summarises external sources, and its output enters the platform only as a clearly labelled input that other agents can cite. Internal data never leaves to the outside world, and outside content never silently mixes with internal records.
Measuring each agent
Because every agent owns one result, every agent gets one primary measure. For inspection and acceptance it might be the time from site visit to signed acceptance. For revenue and payments, the share of invoices paid on time. For exit and continuity, how early risks are flagged before a contract ends. These measures go on the sponsor's dashboard next to the human team's own numbers, so the conversation stays on results rather than on technology.
One practical note on naming: give each agent the name of its outcome, not a persona. "Revenue & Payments" tells a reviewer exactly what the agent is responsible for and what to check. A friendly character name does not, and in a regulated setting clarity is worth more than charm. The same applies to dashboards and logs, where a plain name makes every line easy to trace back to its owner.
Take each agent on your diagram and finish this sentence: this agent is accountable for ___, and we will know it works when ___. If the first blank is a department name rather than a result, slice again. If the second blank is empty, you do not have a use case yet; you have a feature.
