Autonomous agents
already act.
Nobody agreed how.
Elendil is a research initiative and an open community of researchers working out how autonomous AI agents should be governed. We are not a vendor and not a lobby. We take four questions — who is acting, by what authority, who answers for the consequences, which rule actually works — and try to settle them with evidence anyone can check.
Independent
No vendor, ministry, or lab owns our conclusions. Funding is disclosed; findings are not negotiated.
Open by default
Methods, data, code, and the experiments that failed — published, so the work can be checked rather than believed.
Evidence before opinion
If no measurement could change our answer, we call it an opinion and label it as one.
The problem
Deployment compounds. Rules arrive in steps.
Agentic systems are shipped continuously and copied instantly. Governance instruments — registries, thresholds, audits, liability rules — arrive one negotiated step at a time. The distance between the two curves is where the harm lives, and it is widening.
Closing it is not only a legislative problem. Most proposed rules for autonomous agents have never been tested against a system that acts continuously, at scale, and across borders. Some of them cannot be enforced at all. Finding out which ones is research, and that is the work we do.
Research tracks
Four questions, asked properly.
Governing an autonomous agent means answering four things in order. Each is a standing track, with open questions, published evidence, and a working group anyone competent can join.
How we work
From an argument to something you can check.
Agent governance is full of confident positions and thin evidence. Our method is the correction: every claim we make has to be attached to something a stranger could reproduce or refute.
Frame the question
Turn a governance argument into a question that evidence can settle. If no measurement could change our answer, it is an opinion, and we label it as one.
Build the evidence
Assemble datasets, benchmarks, and reproducible experiments — evaluated out-of-time, with uncertainty stated and the failed attempts published alongside the results.
Prototype the instrument
Build the smallest working artefact that shows whether a proposed rule is enforceable at all: a detector, a registry schema, an attestation format, an audit procedure.
Publish & convene
Put the finding in front of regulators, labs, and operators — in the open, in language a non-specialist can check — and revise it in public when someone shows us we are wrong.
Can a mandate be revoked faster than an agent can act on it?
Latest
What we have published.
Working notes, benchmarks, and negative results. Written to be checked, not to persuade.
The community
A think tank is only as good as the people arguing inside it.
Elendil is deliberately plural: machine-learning researchers, lawyers, security engineers, and public-sector practitioners, because no single discipline can answer these questions alone. Working groups form around a track, publish under their own names, and are free to disagree with each other in print.
The initiative is young, and most of the seats are still open — affiliated researchers, working-group members, reviewers, students looking for a thesis subject, and institutions willing to put a hard question to us.
How to take partThe rules for autonomous agents are being written right now.
They should be written by people who have tested them. Bring a question, a dataset, a disagreement, or a chapter you want to work on — and help us settle it in public.