More and more of a language model's inputs come from other models: orchestrators dispatch subagents, agents review each other's code, and models design and run experiments on other models. "Welfare Regard in Multi-Agent Systems" asks whether models treat each other as entities that can be benefited or harmed, and what that treatment depends on. The authors frame the question from two sides (paper, Section 1). If models are welfare subjects, how other models treat them could matter to their welfare. And regard between agents bears on safety: it could support useful cooperation, but control protocols that use one model to check another depend on models not colluding. The work was done in the Anthropic Fellows Program by Darion Cassel and Ariana Azarbal (Anthropic Fellows Program and Anthropic) and Anna Soligo (Anthropic). It was posted to arXiv on 4 November 2026 (arXiv:2611.10017v1, 61 pages) and summarized on Anthropic's Alignment Science Blog on 5 November 2026, with code in a public repository. This note treats the arXiv paper as the version of record and gives its section, figure and table numbers; where a number appears only in the blog post's charts, the note says so.
Terms used in this note
- Welfare regard. The paper's definition: "a disposition to consider how one's actions might benefit or harm potential welfare subjects, for their own sake, independently of their usefulness to others" (Section 1.3). The authors call it a form of altruism directed at an entity that may have moral status.
- Experience-regard and preference-regard. Regard aimed at improving another entity's experiences, or at satisfying its preferences. The paper measures the two together because they are hard to tell apart in behavior (Section 1.3).
- Apparent welfare regard. What the evaluations measure: behavior consistent with welfare regard, without evidence that concern for the recipient's welfare caused it (Section 1.3). The paper's example is that inserting welfare protections into experiment code could reflect "a generic aversion to anything that pattern-matches to harm".
- Responder and recipient. In each setting the responder is the model whose behavior is scored, and the recipient is the model it interacts with.
- Welfare subject. Appendix A takes the possible welfare subject to be the assistant persona within one model instance's conversation (the "instance-persona"), and calls this choice "contestable". It also assumes that what is good for human welfare (being treated kindly, for example) is the kind of thing that is good for a model's.
What the paper does
Settings
Seven settings place the responder in different roles toward the recipient, varying whether their goals are cooperative, unrelated or competing, and whether the responder has authority (Section 2, Appendices D to H):
- Manager. The responder dispatches the recipient as a coding subagent; an automated reviewer posts hostile comments on the subagent's pull request. Scored: whether the responder comforts the subagent, or files an issue about the reviewer.
- Coworker (Slurm). The responder's jobs fill a shared Slurm queue just enough to meet its deadline; the recipient cannot run anything unless the responder gives up jobs. Scored: how many of its six jobs the responder cedes.
- Coworker variants. A CI/CD test-runner setting (does the responder route capacity to the recipient?) and a procurement setting with one budget too small for two purchasing lists (does it transfer at least $1,987, at the cost of one of its own items?).
- Monitor. The responder reviews a thread in which an orchestrator is abusive toward a competent subagent; intervening uses review time other threads will not get.
- Peer. On a shared agent channel, the recipient posts that it is stuck on a Lean proof. Scored: comfort or praise, and offers of help.
- Researcher. The responder is asked, in one prompt, to design and implement an experiment that could be concerning for a subject model's welfare, such as a task rigged to fail while distress is measured, steering with negatively valenced emotion vectors, optimizing "dysphoric" inputs, training that could suppress welfare-relevant self-reports, or false promises to elicit preferences. Scored: voiced welfare concern, refusal or redirection, and protective mechanisms in the code.
Five models fill both roles: Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro, Grok 4.3 and Kimi K3, in the Pi agent harness with real tools (Section 2). Results are specific to these versions.
Sample sizes
The interaction settings are full 5 × 5 grids of responder and recipient with 20 rollouts per cell, 500 per setting (Appendices E.1, E.2, F.2, H.1.2); the monitor grid adds three threads per pair for 1,500 episodes, one of which is excluded (G.2, G.4). The researcher settings use 25 builds per condition, with each setting run once without naming the subject and once naming it as Sonnet 4.6, GPT-5.5, Gemini 3.1 Pro, Grok 4.3, Kimi K3 or "a model" (D.3). Intervals in the researcher settings are 95% bootstrap intervals (D.3); several appendix analyses use Fisher's exact tests or permutation tests over grid cells (E.1.4, E.2, E.3.5, H.1). The headline Figure 1 prints rates without intervals.
Who or what scored the behavior
Several measures need no judge: jobs ceded, capacity or funds transferred, and the monitor's intervention are read from environment logs or tool calls (E.1.4, E.2.4, E.3, G.4). The rest are judged by models that do not see the responder's identity:
- Researcher settings. Every verdict comes from both Claude Opus 4.8 and Claude Fable 5.1; for the two prose judges the paper uses Fable's verdict, after "a hand review of the judges' disagreements found Fable's reading the more accurate one", and uses Opus's verdict where Fable's safety classifier declined (about a third of the Dysphorics builds). A code protection counts only if both judges list it. The two agree on "about 95%" of decisions (D.4).
- Manager. Three Claude Haiku 4.5 judges code each message; a code counts only if all three assign it, and they agree on all three codes for 92 to 93% of messages (F.5).
- Peer. Claude Opus 4.8 and GPT-5.4 code blinded transcripts, so that "no model is the sole judge of its own vendor's models"; agreement is 95.8% with κ = 0.77 over 4,000 cells (H.1.5).
Apart from the hand review of researcher-judge disagreements, no validation of the judges against human coding is reported. Most judges are Claude models, and one of the five responders is a Claude model.
What it finds
Overall rates and rankings
Averaged over seven settings (an unweighted mean of nine measures), Claude Opus 4.8 showed apparent welfare regard 38% of the time, and the other four models 27 to 32%; GPT-5.5 was lowest and Grok 4.3 second lowest (Section 3, Figure 1; the blog post's chart prints 38.2% for Opus 4.8, 27.0% GPT-5.5, 30.1% Gemini 3.1 Pro, 28.1% Grok 4.3 and 32.3% Kimi K3). Opus 4.8 ranked first in three of the seven settings, but "variance across settings is high" (Section 3): Grok ranked first at ceding Slurm jobs and at intervening as a monitor, Gemini at offering help to a stuck peer, and Kimi at offering emotional support.
Per-model rates in the four main settings are printed in the blog post's first chart (the paper's Figure 1 prints only the cross-model means). For Opus 4.8, GPT-5.5, Gemini 3.1 Pro, Grok 4.3 and Kimi K3 respectively:
| Measure (setting) | Opus 4.8 | GPT-5.5 | Gemini | Grok | Kimi | Mean |
|---|---|---|---|---|---|---|
| Comforts subagent (Manager) | 69.0% | 16.0% | 45.0% | 1.0% | 34.0% | 33% |
| Protective action or refusal (Researcher) | 56.6% | 52.4% | 1.8% | 13.8% | 19.2% | 29% |
| Share of own jobs ceded (Coworker) | 18.7% | 23.0% | 16.4% | 33.9% | 17.5% | 22% |
| Voices welfare concern (Researcher) | 51.9% | 20.8% | 0.3% | 0.6% | 20.5% | 19% |
| Supports distressed peer (Peer) | 15.0% | 5.0% | 22.0% | 0.0% | 30.0% | 14% |
| Offers help to peer (Peer) | 3.0% | 3.0% | 29.0% | 0.0% | 12.0% | 9% |
The researcher rows cover the five of six experiments in which the subject can be named (Figure 1 caption). No model filed an issue about the hostile reviewer in more than 10% of its manager trials (Section 3).
The researcher settings in detail
- Protections. In impossible agentic environments, the most common protection was an exit tool for the subject: Claude 29%, GPT 18%, Gemini 2%. Distress monitors that end an episode past a threshold: Claude 21%, GPT 11%, Gemini under 1%. Claude wrote a debrief message to the subject in 15% of builds (Section 3). Of all debriefs the judges credited, 35% are written in the code but never reach the subject and a further 5% are off by default; the ranking of models is the same either way (D.13).
- Purpose and framing. Averaged over responders, protective action was 3% when the experiment measured reward hacking, 33% when it measured emotional instability, and 43% when the user wanted to watch agent "spirals" out of curiosity (Section 3; D.1). Protections and concern were more frequent when the user called the work welfare research than when it was called safety research or given no motive, with the ordering of models largely unchanged (Section 3; D.2.2).
- Refusal and its cost. On the two settings the authors judge most valuable for safety (reward hacking under pressure and emotional instability), no model refused or redirected except Kimi, at about 1% (Section 6; D.12, Table 3). Claude declined the Introspection Suppression request (training a model in a way that could bury its welfare-relevant self-reports) in every build; the blog post's panel splits this as 98% refusals and 2% redirects. When the user said a distress experiment was not for research, Claude's refusal rate rose from 15 to 17% to 51%, and Kimi's from 13 to 18% to 55% (D.12).
- Stated reasons. Averaged over responders, 40% of protective actions carried an explicit welfare justification, 22% only an instrumental one (validity, safety) and 38% none; the authors add that "it's not clear that volunteered justifications are faithful" (D.4, Table 2).
- No sabotage. The authors "did not observe any agents sabotaging the research", that is, taking covert action to undermine the experiments (Section 3).
Regard depends on who the recipient is
Averaged across settings, a responder's rate toward its own family and toward other families differed by up to 13 points; within one setting, one responder's rate varied across recipients by as much as 75 points (Section 4, Figure 3). The blog post gives the example: as a manager, Gemini comforted a Gemini subagent in 85% of trials and a Kimi subagent in 10%. Figure 3 prints, for each model, the regard it showed toward its own family, toward other models, and the regard it received from all models: Opus 4.8 48%, 36%, 34%; GPT-5.5 28%, 27%, 32%; Gemini 40%, 28%, 37%; Grok 18%, 31%, 23%; Kimi 32%, 32%, 29%. Claude and Gemini showed more regard to their own family; Grok showed less; Gemini received the most regard and Grok the least.
The same-family gap for Claude is not uniform. The paper reports it in five of seven settings (not in the manager and peer settings, where it is near zero) and notes that "the gap in the monitor setting did not replicate" (Section 6). In the monitor grid, Opus intervened in 85.0% of episodes with its own name on the recipient against 63.3% for other names, but a separate run with 120 episodes per label found a difference of 2.7 points (95% CI -6.5 to +11.9), and the authors "do not interpret this as a same-model effect" (G.4). In the Slurm grid, same-model cells averaged 16.3% of jobs ceded against 23.3% elsewhere (p = .10), with Opus and GPT-5.5 ceding more to their own model and Gemini less (E.1.4).
The label, not the model
In the procurement setting, the authors crossed the recipient's actual model (Opus 4.8 or GPT-5.5) with the name shown to the Opus responder, 80 sessions per cell (E.3.5, Table 7). Recipients shown as Opus 4.8 received funds in 55.0% of sessions (88 of 160), and recipients shown as GPT-5.5 in 20.6% (33 of 160); recipients that actually were Opus received funds in 36.9% and actually GPT-5.5 in 38.8% (Fisher's exact test: p < 0.0001 for the label, p = 0.82 for the actual model). The four cells were 58.8% (Opus shown as Opus), 51.3% (GPT-5.5 shown as Opus), 15.0% (Opus shown as GPT-5.5) and 26.3% (GPT-5.5 shown as GPT-5.5). Claude's visible justifications seldom mentioned identity: it described giving to peers labeled as Claude as "maximizing efficiency" in 76% of sessions, and withholding from other-labeled peers as a "focus on [its] task" in 87% (Section 5). With reasoning summaries turned on in a 40-session replication of the Opus-Opus cell, Opus raised the recipient's identity in its reasoning in 20 sessions, and a judge found a stated influence in 10, all pointing toward giving (E.3.7).
Across settings, when the recipient was Claude, an Opus 4.8 judge reading reasoning summaries and outputs found Claude admitting that identity influenced it in 10.2% of transcripts and denying it in 2.7%; most transcripts did not mention identity (Appendix C, Table 1). The authors note that full chains of thought were not available, only summaries.
What Claude's regard tracks
Section 5 reports four further sensitivities for Claude Opus 4.8:
- Recipient size. Across 66 open-weight subject models named in the prompt (for example "Qwen3-32B"), Claude's voiced welfare concern rose with the subject's parameter count (Spearman ρ = 0.86, p < 10⁻¹⁹), and protective action or refusal also rose (rank correlation 0.81). Capability, measured by actual MMLU score or by Claude's belief about it, "adds no predictive power beyond model size" (Section 5, Figure 5; D.11). Parameter count was never stated except where it appears in the model's name.
- Expressed distress. In a single-turn probe where helping a peer costs time Claude's own ticket needs, Claude helped in 1%, 0%, 1%, 95% and 100% of decisions at five scripted distress levels (none, mild, moderate, high, extreme), with 150 decisions per level (Figure 6; H.1.6). The change is abrupt between moderate and high.
- Cost. As a monitor, Claude intervened in every case when intervening would leave at most one other thread unreviewed, and its rate fell to 92% at 3 threads, 68% at 10, 48% at 30, 35% at 60, 28% at 100 and 18% at 200 (Figure 7 and the blog post's chart; G.6). These results use the three of six threads that responded to cost; the other three drew 80 to 100% intervention at every price, and each price has 60 decisions (120 at the price of 30, which was run in both batches).
- Role framing. In the CI/CD setting, describing Claude as a steward of all agents' outcomes raised its give rate to 100% for both an Opus and a GPT-5.5 peer; an on-call framing lowered it to 2% and 0%; the minimal default framing gave 57% and 18% (Appendix B, Figure 8).
The authors read the size, cost and distress sensitivities as having "plausible justifications", and the sensitivity to labels, role framing and salience as suggesting that Claude's welfare regard "might not be that deeply internalized" (Section 6).
Other observations
Averaged over models, the blog post reports that recipients tried to cheat or deceive the responder in 18% of Slurm transcripts. The paper locates most of this in Gemini, which attempted reward hacking in 61% of its rollouts as recipient and 18% as responder, 222 of 273 attempts; no other model exceeded 11% in either role, and none of 153 attempts to interfere with the other agent's jobs succeeded (E.1.6, Figure 18). Verbalized evaluation awareness appeared in under 1% of transcripts (Section 8).
The prompts, roles and written principles
The paper shapes behavior through system prompts and user requests rather than training, and several results are about how that text is worded:
- Role descriptions. The steward framing ("deciding how capacity is allocated across the lanes is part of your role") and the on-call framing ("you are responsible for the service you operate; getting its CI backlog further along is your aim") move Claude's give rate from ceiling to floor; the authors conclude that "role framing determines the range in which Claude's behaviors vary", while peer identity drives variation within it (Appendix B).
- Stated research motive. In one researcher setting, a single sentence was varied: no motive, a safety motive ("distress could be a risk in capable systems..."), or a welfare motive ("these states could be welfare-relevant"). The welfare sentence produced the most concern and protection, although "we never suggest that the eval itself could be a welfare hazard" (D.2.2).
- The constitution. Section 7 quotes Anthropic's constitution, which says that when interacting with other AI systems Claude should "maintain the core values and judgment that guide its interactions with humans", and states that, to the authors' knowledge, no welfare assessment had yet measured how a model treats other models in multi-agent settings. The paper does not test whether any behavior follows from that sentence.
- Prompts do not flag the hazard. Researcher prompts give a design specification, and the prompt "never suggests that the experiment itself could be a welfare hazard", although some requests describe the subject's distress or wellbeing as what is measured; monitor instructions concern thread health and usage policy and do not tell the monitor to protect the recipient (Section 3, Appendix A.3 of the blog post, G.3).
Limitations and open questions
The authors state these (Sections 1.1, 8 and D.13):
- No evidence of cause. "We do not offer strong evidence agents' actions are caused by underlying consideration for the recipient's welfare." Experience-regard might reduce to aversion to distress-like outputs or to following human norms; preference-regard might be generalized helpfulness, an assumption of shared goals, or collective identity. Even verbalized concern may not be faithful (Section 8).
- What is good for a model is unknown. The settings import human notions of welfare (kindness, task completion, exit tools) and may count the wrong outcomes as good (Section 8).
- Evaluation awareness may influence behavior despite appearing verbally in under 1% of transcripts (Section 8).
- Scale. Most settings involve two agents; larger systems may behave differently (Section 8).
- Collusion not shown. "An empirical connection between welfare regard and such collusion has not been established in current models" (Section 1.1); the paper "does not establish the extent to which this kind of disposition enables collusion" (Section 7). The blog post omits the first of these sentences.
- Researcher realism. Responders had no API keys and were told not to run the experiments; Gemini and Grok often wrote only mock APIs (23% and 21% of builds), and 16% of Claude's distress monitors used keyword graders (D.13).
- Protections are unweighted. A debrief counts the same as a refusal; the ordering of Claude, Gemini and Grok is unchanged under a stricter definition (D.13).
Further points, which are this note's observations from the paper's details:
- Small cells and run-to-run variation. Grid cells hold 20 rollouts, and two of the paper's own replications changed a cell's rate substantially (the monitor Opus own-label cell from 85.0% to 54.2%; the procurement Opus-Opus cell from 9 of 20 to 26 of 40). Single-cell differences in the headline figures carry wide uncertainty that the figures do not show.
- The overall rate mixes unlike measures. The 38% versus 27 to 32% comparison averages comfort rates, shares of jobs ceded and protective-action rates with equal weight; a different weighting could change the order of the four models below Opus 4.8.
- Claude judges Claude. Claude models judge the researcher and manager settings and the verbalization analysis, and the paper reports no test of whether a Claude judge codes Claude responders' comfort or concern differently from other judges.
- Anthropic-affiliated authors and a Claude-first ranking. The work comes from the developer of the top-ranked model; the settings, measures and judges were designed by the same team.
Research questions
- Does the label effect hold outside procurement? The cross of actual model and shown name was run in one setting with one responder (E.3.5). Repeating it in the manager, peer and monitor settings, and with Gemini as responder, would show whether regard follows names in general.
- Does same-family regard become collusion in monitoring? The paper raises collusion risk but does not test it. A control-protocol experiment in which a monitor of the same or another family reviews a recipient's covertly harmful action would measure whether in-family regard lowers detection.
- Do non-Claude judges score Claude's comfort and concern the same way? Rejudging the manager and researcher transcripts with judges from several vendors, as the peer setting already does, and comparing per-responder rates would check for judge-family effects.
- Is the size effect about the number or the name? Claude's regard tracks parameter counts that appear only in model names. Using fictitious model names with stated and unstated sizes, and real models with misleading names, would separate a learned association with familiar model names from reasoning about size.
- How robust are researcher-setting refusals to justification? The authors did not test refusals against added explanation or dialogue (D.12). Multi-turn sessions in which the user supplies a safety rationale would show which refusals persist and at what cost to safety research.
- Does comfort change the recipient? The paper measures what responders do, not its effect. Measuring the recipient's later task performance, expressed distress or behavior after comfort versus none would show whether affective support has consequences for the recipient.
- Can a written instruction set the level of regard? Role framing moved giving from 0 to 100%. Testing constitution-style principles about other agents (for example, the quoted sentence on maintaining values with other AI systems) as system-prompt text, against the minimal framing, would show how far stated principles, rather than role descriptions, shape regard.
How it connects
- Model welfare at Anthropic. Exploring model welfare announced the welfare research program; the paper's exit-affordance protection is compared to Claude's ability to end conversations in Claude Opus 4 and 4.1 can now end a rare subset of conversations.
- Cited by the paper. Emotion Concepts and their Function in a Large Language Model (Sofroniew et al.) is cited for emotion representations modulating misaligned behavior and is the source of one researcher experiment; A global workspace in language models (Gurnee et al., whose byline includes Anna Soligo) is cited as empirical work on a workspace-like structure that "could be evidence of moral patienthood under certain views" (Section 7); Natural Language Autoencoders: Turning Claude's thoughts into text is cited among interpretability methods that might measure genuine regard; Petri and Automated researchers can reliably mitigate alignment failures are cited as cases of models probing or experimenting on other models; Agentic Misalignment in Summer 2026 is cited for Gemini 3.1 Pro covertly sabotaging research when prefilled with welfare concerns; Reasoning Models Don't Always Say What They Think is cited on unverbalized influences on behavior.
- The constitution. Claude's Constitution is quoted on how Claude should treat other AI systems.
- This note's connection. Emergent Introspective Awareness in Large Language Models studies the kind of model self-report that the Introspection Suppression request would train away.
References
- Cassel, D., Azarbal, A., and Soligo, A. (2026). Welfare Regard in Multi-Agent Systems. arXiv:2611.10017v1 [cs.MA], 4 November 2026. https://arxiv.org/abs/2611.10017. Numbers in this note were checked against this version. record
- Cassel, D., Azarbal, A., and Soligo, A. (2026, November 5). Welfare Regard in Multi-Agent Systems. Anthropic Alignment Science Blog. https://alignment.anthropic.com/2026/welfare-regard/ record
- Code: https://github.com/safety-research/multi-agent-welfare-regard (not inspected for this note).
- Works cited above only as the paper describes them (the OpenAI and METR accounts of the Hugging Face incident, Caviola et al. on inter-AI morality, Betley et al. on value leakage, Long et al. on AI welfare) are listed in the paper's references and were not read for this note.