AI Alliancen / Weekly newsletter
Week 40 202628 Sept – 4 Oct 2026
From AI experiments to responsible operation
What may an agent do? Who reviews its work? And how do we know whether AI helps? This week brings together three current developments, three outside perspectives and a concept connecting them.
You are reading an earlier edition. Dates and information reflect that edition. Read the latest
This week’s common thread is practical control: clear tasks, limited access and people who can intervene. Here is our selection as of 2 October, distinguishing research, guidance and rules. The perspectives section also includes older, useful knowledge with its original source dates.
The most important developments
The week’s news
Perspectives to take into your work
Beyond the network
Make room in the calendar
Coming up
PwC Denmark · Online
How can AI scale across the public sector?
External webinar · free according to the organiser. Check registration terms.
KL · Online
AI and communication with citizens
External webinar · free. For municipal staff and managers.
AI Alliancen · KL-huset · Copenhagen
The next AI Alliancen meeting
Network meeting for members and invited guests. No public registration.
Times are Danish local time. External events are organised by the named providers. Check their current programme and terms. See the full calendar
The week’s concept / Explained properly
What is an agent harness?
An agent harness is the software around an AI model that lets it work as an agent: keeping track of a task, using tools and continuing across multiple steps.
Think of the model as the part proposing the next step. The harness helps move the work forward: What information does the model need? Which tools are available? What happened last time? What happens when a tool fails?
There is no single standard definition of its scope. A harness may focus on developer productivity or include more governance and security functions. The name alone does not tell you how secure a system is.
A simplified workflow
Agent harness · software around the model
- 02Model proposes
- 03System checks access
- 04Permitted tool runs
Context and progress
Keeps task state, relevant information and intermediate results so work can continue without starting again.
Tools and access
Connects the model to search or a document store. The system must enforce access appropriate to the task.
Controls and approval
May require human approval before consequential actions. An instruction in a prompt is not a technical barrier.
Logging and stop
May record actions and errors and let an operator stop a run. This needs actual implementation and testing.
Illustrative municipal example
An illustrative example: drafting a reply to a citizen
A staff member wants a draft about waste sorting, based on the municipality’s public guidance. The agent gets a bounded task, not unrestricted access to municipal systems.
- The staff member provides the question and approved source material.
- The agent searches only the permitted guidance. System permissions prevent access to other registers.
- The model drafts a reply with sources. If evidence is missing, the workflow must be able to stop and ask for help.
- The staff member checks the content and sources. The agent has no access to send the reply to the citizen.
A harness is a framework, not a guarantee. Incorrect answers, a missing legal basis or excessive permissions do not disappear because a system has a harness. Controls need testing and the organisation remains accountable.
Sources behind the explanation
Our explanation and illustrative example draw on the sources below. Anthropic describes one implementation; the security sources inform the control principles.
Editorial selection and explanation: AI Alliancen · Prepared with AI assistance and source checks. Public-sector applications are editorial suggestions, not measured effects or legal approval.