Manifesto
Govern what agents do, not what they are.
We started the Elendil Initiative because the rules for autonomous agents are being drafted faster than anyone is testing them. This is the position we argue from, and the four commitments we hold ourselves to while arguing it.
The problem
The shortage is not opinion. It is evidence.
Autonomous agents are already deployed inside enterprises, public administrations, and markets. They act continuously, they are copied instantly, and they cross borders without noticing them. Meanwhile the governance conversation is loud, confident, and almost entirely unmeasured.
People propose agent registries without knowing whether an unregistered agent can be detected. They propose revocable mandates without knowing how long revocation takes. They propose capability thresholds for actors that can be divided in half at zero cost. These are empirical questions, and answering them is research, not advocacy.
A rule that cannot be enforced is not a safeguard — it is a promise made to the public on someone else's behalf. Before Europe writes more of them, someone should find out which ones work.
“An agent should be governed by the authority it was granted, the record it leaves, and the harm it can reach — not by what it is made of.”
Model-centric governance ages badly: capability thresholds move, architectures change, and the same system behaves differently depending on what it is wired to. Mandate, record, and reach are properties of the deployment, they are observable from outside, and they are the things an affected person actually needs to establish. That is a position, not a finding — and we hold it only as long as the evidence supports it.
Our commitments
Four commitments, written to be enforced against us.
Each one is a constraint on what we accept, what we publish, and who we work with. They are specific on purpose: a commitment you cannot be caught breaking is decoration.
Independence
No vendor, ministry, lab, or donor owns our conclusions. Funding is disclosed on the face of each note, members declare their interests before they start work rather than after, and we decline any arrangement that would give a funder influence over a finding, its timing, or its release. We would rather publish less than publish captured.
Open by default
Methods, data, code, and the experiments that failed are published together, so the work can be checked instead of believed. Where data cannot be released — because it is personal, or because a partner has a legitimate constraint — we say precisely what is withheld and why, and the claim that depends on it is marked as unverifiable by outsiders.
Evidence before opinion
A claim earns a place in our work only if we can say what would refute it. Numbers arrive with uncertainty, out-of-time evaluation, and their limits stated. Where we hold a view that no measurement could settle, we label it an opinion and let the reader discount it accordingly — which is most of what passes for agent-governance debate today.
Held to our own standard
We are asking that autonomous systems keep records, declare their authority, and answer for their effects. An initiative that asks for that owes the same: named authorship, named reviewers, dated corrections that leave the original readable, and a public account of what we chose not to publish. Hold us to it, in writing, and we will answer in writing.
What we are building toward
What this looks like if it works.
A shared vocabulary between engineers and regulators
Mandate, revocation, attribution, and record are currently used to mean four different things in four different rooms. Precise, agreed definitions are unglamorous work and the precondition for everything else.
Rules that were tested before they were passed
A world where proposing an instrument for autonomous agents comes with an obligation to show it can be enforced — and where the enforceability evidence is public, reproducible, and older than the political deadline.
An evidence base that outlives the initiative
Benchmarks, datasets, and reference implementations that belong to the field rather than to us. If Elendil disappears and the evidence base remains useful, that is a success, not a loss.
Disagreement that stays in public
These questions are political as well as technical, and consensus reached too early is just the loudest party winning. We would rather convene the disagreement, publish both sides with their evidence, and let it be resolved where everyone can see it.
If you think this position is wrong, that is exactly the conversation we want.
Bring the argument, and bring what would settle it. Disagreement with evidence attached is the most useful thing anyone can send us.
Take part