Agentic AI: Development, Use & Governance – Full programme

Session details

Part 1 – Foundations: what is an agent, and what is agency? (14:05–14:30)

The opening session gives the room a shared vocabulary: the technical picture, the philosophical one, and the economic one. The three talks run technical → conceptual → institutional, so each adds a layer to the last.

1.Benjamin Ricaud (UiT) –Introduction to Agentic AI. Agents, harnesses, tokens, vibe coding, current technical challenges and the next six months.

2.Joey Pollock (UiO, IFIKK) –What is agency?Contemporary philosophy of agency and what it implies for accountability and anthropomorphism.

3.Andreas Føllesdal (UiO) –Humans and agentic AI as a principal–agent relationship: is anything new?Agency drift, agency costs, and "fractal" agency risks when principals delegate to AI.

Discussion prompt: do the technical, philosophical and principal–agent accounts of "agency" point at the same thing – and does it matter for governance if they do not?

Part 2 – Use: agentic AI in research practice (14:30–14:55)

Three first-hand accounts of running agentic systems in research at UiO, moving from the normative question (what should humans contribute?) to a worked example (a database built by an agent) to the practical stack.

1.Geir Kjetil Sandve (UiO, IFI) –Human contribution in the age of agentic AI. What humans should conceive, decide, check and vouch for – and why we should declare "human use" rather than "AI use". Includes an invitation to a joint manuscript.

2.Bjørn Høyland (UiO, ISV) –Building a traceable database of European Parliament politics with an AI agent. StREP: almost all code written by a coding agent since April 2026; how to stand behind data without having read the code.

3.Lex Nederbragt (UiO) –Do you really need a Claude Code subscription when working at UiO?University-provided resources and where they sit in the agentic software stack.

Discussion prompt: Høyland's and Sandve's talks are two answers to the same question – what does it take to vouch for work you did not do? Nederbragt's talk invites the practical follow-up of what UiO should provide.

Part 3 – Development: building agents you can trust (14:55–15:30)

The SINTEF block: three engineering approaches to the same problem – agents produce output that looks right and is wrong – using verification gates, structured pipelines and contracts. Placed last in the first block because Shivashankar is only available from 14:50; the block's discussion follows directly and links to governance.

1.Karthik Shivashankar (SINTEF) –Deterministic gates for agentic coding. Live demo of Qloop: LLM-free gates (static analysis, contracts, syntax, tests, complexity) behind the agent, with failure reasons fed back. Allow 10 minutes for the live demo.

2.Gencer Erdogan (SINTEF) –From manual to automated cyber risk assessment. The open-source CORAS Threat Modeler: an LLM- and RAG-driven multi-agent pipeline from system descriptions to threat models and mitigations.

3.Adela Nedisan Videsjorden (SINTEF) –Agents for compliance-by-design data/AI pipelines and contract-based interactions. Why direct LLM generation of pipelines fails, and how agents plus contracts make them compliant and executable.

Discussion prompt: where should an LLM not be trusted in a pipeline, and are deterministic gates and contracts a general governance pattern or a coding-specific one? This links directly to Session 4.

Governing Agentic AI: from the AI Act to an international agency (15:45–16:30)

Four talks that scale up from an experiment to the EU framework, to liability, to the international level. Ricaud's second talk opens the session with empirical results, so the legal talks respond to evidence rather than abstractions.

1.Benjamin Ricaud (UiT) –Can AI agents make law?Multi-agent debate experiments on autonomously drafting a hypothetical constitution for AI agents.

2.Tobias Mahler (UiO) –Agentic AI and the AI Act. Does the EU AI Act sufficiently address the risks of agentic systems?

3.Cláudia Lima Costa (University of Porto) –Agentic AI: do we need compensation funds?When fault and causation become unprovable, should compensation schemes replace traditional liability?

4.Malcolm Langford (UiO/TRUST) –Governing agentic AI – internationally. Concrete proposals for the Norwegian–Finnish initiative on common standards and a new international agency for safe AI, and how to design a process that survives political disagreement.

Discussion prompt: three governance levels (EU, national liability, international agency) – which one is the binding constraint for agentic AI, and what from Sessions 1–3 should each of them take on board?

Discussion (16:30–17:00)

Chaired round on the workshop's stated aim – mapping who is doing what and sparking collaboration.