The Big Picture
Agents trained to both probe a partner’s abilities and actively steer behavior coordinate better with teammates that can adapt, finding higher-value joint strategies instead of accepting the first workable option.
ON THIS PAGE
Key Findings
Training a generalist agent against a structured hierarchy of partners that includes capability-limited, adaptive followers causes the agent to learn when to probe, when to follow, and when to lead. That behavior—persistently signaling higher-value conventions when a partner is capable of switching—lets teams converge on better shared solutions than agents trained only to react. The approach matches prior methods when partners are fixed but outperforms them in settings where partners can adapt to a restricted set of conventions. Consensus-Based Decision Pattern
Test your agentsValidate against real scenarios
By the Numbers
1Two evaluation domains: a repeated matrix coordination game and a point-mass rendezvous task (the rendezvous layout uses 4 landmarks).
2Training follows a 3-stage hierarchy (base conventions K0, restricted adaptive followers K1, and the final generalist K2).
3Matches baselines on uniform-payoff games and with fixed (K0) partners, but shows notably higher coordination efficiency when paired with adaptive K1 partners in differentiated-payoff games.
Why It Matters
Engineers building multi-agent systems and teams of AI agents should care because agents that can probe and steer teammates will reach better joint outcomes with unfamiliar but adaptable partners. Technical leads evaluating agent-to-agent evaluation or multi-agent trust can adopt this training hierarchy to improve real-world team reliability where teammates may have limited but nonzero flexibility. Evaluation-Driven Development (EDDOps)
Key Figures
Fig 1: Figure 1. Comparison of the ConventionPlay pipeline with existing ad-hoc collaboration methods. While baseline approaches diversify the K 0 K_{0} population, we introduce diversity among K 1 K_{1} followers to force the K 2 K_{2} agent to infer its partner’s limited repertoire and coordinate on the most effective shared convention.
Fig 2: Figure 2. Payoff matrices for the Repeated Matrix Game. Left: Uniform payoffs across conventions. Right: Differentiated payoffs across conventions.
Fig 3: Figure 3. Trajectories of agents playing two variants of the point-mass-rendezvous game. Left: Layout with 4 landmarks with equal value; no reward is achieved as agents navigated to different landmarks. Right: 4 landmarks with differentiated rewards, visualized by opacity; both agents successfully navigating to the landmark with the highest rewards.
Fig 4: Figure 4. Makeup of the generated K 1 K_{1} subsets. The top left image has cross-play of all the K 0 K_{0} population; clusters here roughly visualize compatible policies which can be considered conventions. The top middle shows how the stratified sampling has placed different agents in different clusters. The top right shows the maximum capability distribution across subsets, ensuring various K 1 K_{1} subsets have different capabilities. Bottom left shows Jaccard similarity between all generated clusters. Bottom right shows enumerated subset sizes and the number of times a specific agent is used in the support of the entire K 1 K_{1} population.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
The method depends on how diverse and well-separated the initial set of base conventions is; if the base population does not expose distinct strategies, the benefit shrinks. The capability-sampling step currently uses self-play return to select partners, which can fail to distinguish tactically different but similarly scoring behaviors. Experiments are limited to fully cooperative tasks—extensions to mixed-motive settings will require additional safeguards and evaluation. Capability Elicitation
Deep Dive
Train agents in three stages: first create a diverse set of base conventions (K0) by training many self-play policies; then build a population of restricted, adaptive followers (K1) where each follower adapts only to a subset of the K0 conventions; finally train the generalist agent (K2) against this hierarchy. Exposing the generalist to K1 partners forces it to go beyond blind adaptation: it must probe to reveal what a partner can or cannot change, and decide whether to conform or to persistently signal a higher-value strategy to steer the partner. A2A Protocol Pattern Tested on two settings—a repeated matrix coordination game (with uniform and differentiated payoffs) and a point-mass rendezvous task with four landmarks—this training produces agents that match baseline performance when partners are fixed, yet outperform baselines when partners have limited adaptation capability. The learned behavior is concrete: instead of immediately matching a partner’s early, possibly suboptimal choice, the agent continues to try higher-value options when it detects partner flexibility, enabling the team to converge on the best mutually achievable convention. Main limitations include reliance on return-based sampling to form K1 groups and evaluation only in fully cooperative environments; improving diversity metrics and testing mixed incentives are natural next steps. Red Teaming Pattern
Test your agentsValidate against real scenarios
Credibility Assessment:
Affiliations not specified and very low author h-indices (≈1); arXiv preprint with no citations—insufficient reputation signals (emerging/limited).