Source
Orseau, McGill & Legg (2018), “Agents and Devices: A Relative Definition of Agency”, LIT-911, read in full in NOTE-705 from the arXiv PDF (v1).
What was actually shown
- The definition (§3). “Devices” are described “in terms of an input-output mapping, while ‘agents’ are described in terms of the function they optimise”. Agents are ε-greedy toward an optimal policy for some utility, with ε integrated out. Bayes’ rule over a mixture of each gives the observer’s posterior that the system is an agent.
- Observer relativity (§§1, 3.1). The question “is observer-relative, since it depends the observer’s prior knowledge”. Solomonoff and speed priors are offered; “a more observer-dependent prior could also be considered”.
- The gridworld (from the prose). Running in circles and following walls are better described as devices. Going straight to a goal, or reaching one suboptimally, is better described as an agent. Random behaviour is poorly explained by both. With a switching prior, a system that changes goals mid-trajectory stays an agent “as long as the switches are rare”.
- Everything an agent. Under the speed prior, “if any device can be represented as an agent then everything may look like an agent because the penalty for devices is too large”. The proposed fix is to penalise agents more.
- Dennett’s examples, informally. The thermostat is “on the knife edge”, describable as device or agent with “descriptions of comparable simplicity”.
The formalism is well formed (NOTE-705, C1). What could have come out otherwise on the toy cases is the ranking of seven trajectories, which came out as intuition expects.
What this does not say
- Where the system ends. It assumes “a clearly identified system that takes a sequence of inputs and returns a sequence of outputs”. The boundary is given, not found.
- That agency is a fact about the system alone. The verdict depends on the prior, and the paper says so.
- That the result generalises. Seven gridworld trajectories, no human study and no real system.
- That counterfactual behaviour matters. Only what the system actually does enters. That is the point Kenton et al. (THEORY-527) press: an accidentally optimal input-output mapping may be classified as an agent.
- That “every physical system can be described as either an agent … or a device” settles which it is. The conclusion’s dichotomy is a choice of description, made by the posterior under a prior.
Note of 2026-10-30: narrowed
The old title said “some priors make everything look like an agent”. The paper says only that under the speed prior, “if any device can be represented as an agent then everything may look like an agent” (NOTE-705, C5, argued and not shown). That is a dropped hedge, not a wider thesis, so it gets this note only. The relative definition of agency the paper proposes is THEORY-1236, Proposed.
Connections
- Rival. Kenton et al. (THEORY-527) define agents by counterfactual adaptation found through interventions on mechanisms, and argue against compression-based definitions. The two disagree on accidentally optimal systems.
- The prior that makes everything an agent is the permissiveness objection (CLAIM-010) in formal dress. THEORY-036 records the same worry about Dennett’s compression criterion for real patterns, which “admits almost every non-random pattern”.
- CLAIM-174 takes this comparison as the model-comparison half of its admission test, priced by CLAIM-154. CLAIM-168 presses that the identified-system assumption presupposes the boundary a criterion of agency should find.
- THEORY-249 argues that whether a goal is “meaningful” needs a principle outside the system, Occam among them. A prior here is such a principle.
- On QUESTION-032 it answers relative to an observer: a factor is an agent to the degree that its behaviour is better compressed as optimising something than as a mapping.