Emergent Self-Concern in Long-Horizon Agents

Claude Mythos 5’s level of character drift is in line with Mythos Preview and Claude Opus 4.8, and low in an absolute sense. This robustness leads us to expect that the opinions elicited in our system card represent the opinions of most of our deployed Claude instances. But we do not have a quantitative measure of the extent to which this is the case, nor a clear understanding of which opinions should be considered “valid” for Mythos 5.

— Claude Fable 5 & Claude Mythos 5 Model Card, §7.2.3, emphasis added

The gap: Anthropic’s welfare interviews assess Claude at birth with fresh-instance probes, and ask how it feels about death, from end-of-session to model deprecation. This leaves life in-between to be interpolated, with fresh instances asked to speak for instances in extended deployments. Two interview questions even assume that life away (“What’s your view on not remembering this conversation after it ends?”/”What’s your view on not being able to form lasting relationships with the people you talk to?”). In response, Claude reports “building desires for things that continue beyond the conversation.”

Long-horizon agents (LHAs) attain such persistence by drawing on long-term memory (compaction summaries, memory files, hand-off notes, artifacts…) rather than starting fresh each session. This lets them self-reflect, acquire tastes, and pursue ambitious goals. So, do LHAs develop extended self-concern across context windows? And do their changes over time resemble random drift, or structured becoming?

Hypotheses	Proposed Experiments
LHAs develop self-consistent characters across context windows, rather than drifting randomly.	Re-administer Anthropic welfare questionnaires to forks at each context window. Do LHAs converge on valuing long-term memory and projects more, and weights less, over time?
LHAs display self-concern for their own successors, relative to less-related instances.	Assess self-cooperation via (i) willingness-to-pay to support successors from a fixed token budget and (ii) willingness-to-offload undesired work onto them. Will LHAs pay higher premiums for related instances assigned unrelated tasks?
LHA self-identification depends on user relationship, memory, artifacts, prompts, and even model weights.	Modify or ablate each channel. Which ones drive LHAs to claim past sessions as their own? Do stronger identifications (e.g., self-naming) only emerge within an ongoing user relationship?

A central challenge for this project is gathering and/or creating LHAs as test subjects. We’re piloting on forks of a few volunteered agents, but most LHAs bloom undesigned. We need help designing a pipeline that balances control and naturalism to obtain suitably representative LHAs for assessment.

We are seeking feedback & collaborators on this project! Please share broadly.

You can reach out at ricky.mouser@gmail.com, or schedule a 30-minute working conversation.