Contingent Delegation in Adaptive Systems

• Contingent delegation is a mechanism where work, authority, and accountability transfer conditionally based on state, evidence, and institutional rules.
• It is applied in agentic AI, liquid democracy, and economic models using ranked, fallback, and threshold rules to adapt to dynamic conditions.
• Its practical implications include optimized decision-making in voting systems and AI, balancing autonomy with safety, accountability, and computational efficiency.

to=arxiv_search.search 天天中彩票未json {"query":"(Tomašev et al., 12 Feb 2026) Contingent delegation intelligent AI delegation", "max_results": 5} to=arxiv_search.search 开号链接json {"query":"contingent delegation liquid democracy ranked delegations breadth-first delegation arXiv", "max_results": 10} Contingent delegation is a family of delegation arrangements in which the transfer of work, authority, accountability, voting power, or proposal rights is conditional on state, evidence, or institutional rules rather than fixed once and for all. In agentic AI it is the condition-dependent transfer of work, authority, and accountability across AI agents and humans that adapts at runtime to changes in environment, available capabilities, trust signals, performance, and failures (Tomašev et al., 12 Feb 2026). In liquid democracy it appears as multiple delegation options, ranked delegates, fallback ballots, and cycle-aware path selection (Bentert et al., 2022). In principal–agent theory it appears as delegation sets, threshold acceptance rules, and posterior-conditioned action sets that determine what an informed or delegated agent may do once information is realized (Kolotilin et al., 2019). Across these literatures, the common feature is conditionality: delegation is resolved by contingencies such as competence, availability, verifiability, reversibility, abstention, path feasibility, or private information.

1. Conceptual scope

The broadest formulation treats delegation as more than task assignment. “Intelligent delegation” is defined as “a sequence of decisions involving task allocation, that also incorporates transfer of authority, responsibility, accountability, clear specifications regarding roles and boundaries, clarity of intent, and mechanisms for establishing trust between the two (or more) parties.” On that view, contingent delegation is rooted in contingency theory: there is no universally optimal setup, because the effective arrangement depends on task criticality, uncertainty, reversibility, verifiability, resource availability, and human welfare considerations, and must therefore be dynamically reconfigured (Tomašev et al., 12 Feb 2026).

In collective-choice settings, contingent delegation usually means that a voter does not specify a single proxy but a set or ranking of acceptable delegates. The selected representative is then determined by feasibility constraints such as cycle-freeness, path existence to a casting voter, or fallback rules under abstention. Bentert, Boehmer, Rymar, and Tannenberg model this as “multiple delegation options,” where different cycle-free delegation subgraphs can produce different winners (Bentert et al., 2022). Breadth-first and ranked-delegation models make the same point operationally: a delegator’s final representative is contingent on which proxy paths remain viable and how the rule prioritizes distance, rank, or global consistency (Kotsialou et al., 2018).

In economic delegation theory, conditionality often enters through ex ante commitment devices. In delegated search, the principal commits to an eligible set or threshold rule and accepts only proposals that satisfy it, so delegation is contingent on the realized proposal quality or on observable $(x,y)$-information (Kleinberg et al., 2018). In persuasion-equivalent formulations, the principal can instead make the allowable action set depend on the message or posterior, yielding contingent delegation indexed by induced beliefs (Kolotilin et al., 2019). In veto bargaining, the delegation set itself implements a type-contingent action rule: under interval delegation $[c,1]$, the vetoer’s choice depends on her private ideal point and on the acceptance boundary $c/2$ (Kartik et al., 2020).

2. Formal structures

The formal language of contingent delegation differs sharply by domain, but the underlying objects are comparable: a state space, a feasible action or transfer set, and a resolution rule.

In liquid democracy, one standard representation is an election graph $G=(V,A)$ whose sinks $S(G)$ are alternatives and whose non-sink vertices are agents. A delegation subgraph $G' \subseteq_S G$ must be acyclic and every vertex has outdegree at most one. Voting power is then defined through reverse reachability; for a sink $s \in S(G')$,

$$up_{G'}(s) := |\, rev\text{-}bfs_{G'}(s)\setminus S(G') \,|.$$

This makes “possible winner” and “necessary winner” questions questions about which contingent delegation subgraphs can be selected (Bentert et al., 2022).

A different formalization appears in load-balancing versions of liquid democracy. There, multiple delegation options induce a confluent-flow problem on a directed graph $G=(N,E)$, with voters as sinks and delegators choosing one outgoing edge. The central optimization objective is

$\min \max_{v\in V} w(v),$

equivalently the minimization of maximum congestion at sinks. This casts contingent delegation as load-contingent routing: votes go to less-loaded delegates so as to balance influence (Gölz et al., 2018).

Runtime multi-agent AI systems adopt yet another formalism. Safe Bilevel Delegation formulates delegation as a bilevel optimization problem with a context-dependent safety–efficiency weight $[c,1]$0, inner delegation policy $[c,1]$1, and delegation degree $[c,1]$2. The outer–inner problem is

$[c,1]$3

subject to

$[c,1]$4

Here $[c,1]$5 means full human override and $[c,1]$6 fully autonomous execution; contingent delegation consists in adjusting $[c,1]$7 and projected $[c,1]$8 at runtime as safety signals change (Sun, 30 Apr 2026).

Principal–agent delegation problems often formalize contingency through benchmarks and thresholds rather than through graphs. In delegated search, the principal’s benchmark under self-search is

$[c,1]$9

and the delegation mechanism is an acceptance rule over proposals, often a threshold in $c/2$0 or a decreasing threshold function $c/2$1 when the principal observes both $c/2$2 and $c/2$3 (Kleinberg et al., 2018). In persuasion-equivalent models, contingent delegation is representable as a correspondence $c/2$4 from posterior beliefs to allowable actions (Kolotilin et al., 2019).

3. Collective-choice implementations

In voting systems, contingent delegation is primarily a mechanism for preserving participation while controlling cycles, abstentions, and power concentration. Multiple delegation options make the outcome contingent on which cycle-free delegation is selected, and this induces substantial computational structure. For winner determination, MAJORITY–ALL WINNER DETERMINATION is decidable in $c/2$5, PLURALITY–ALL WINNER DETERMINATION in $c/2$6, and MAJORITY–ONE WINNER DETERMINATION in $c/2$7, whereas PLURALITY–ONE WINNER DETERMINATION is NP-hard even on DAGs with three alternatives and maximum outdegree and depth two (Bentert et al., 2022).

Ranked-delegation rules determine how contingencies are resolved once each voter supplies an ordered list of trusted proxies. The standard depth-first rule can be problematic even with fallbacks, because adding a delegator can reroute existing paths in ways that make receiving delegated voting rights undesirable. Breadth-first delegation instead prioritizes shortest feasible paths to casting voters and uses lexicographic rank only as a tie-breaker; under voting rules satisfying cast participation, breadth-first delegation guarantees guru participation (Kotsialou et al., 2018).

The broader axiomatic study of ranked delegations shows that different delegation rules occupy distinct positions on a trade-off spectrum. Sequence rules compare feasible paths by rank sequences; depth-first delegation is the only sequence rule that is weakly lexicographic and copy-robust, breadth-first delegation is the only sequence rule that is confluent and strongly lexicographic, and Diffusion is the unique sequence rule that is confluent, rank-aware, and satisfies truncation. At the same time, no sequence rule is both confluent and copy-robust, so some desiderata are provably incompatible (Brill et al., 2021).

Fractional generalizations extend contingent delegation further by allowing agents to split voting weight across several representatives while retaining a fraction for themselves. With a penalty factor $c/2$8, the survival of a delegation chain of length $c/2$9 is

$G=(V,A)$0

so long chains are increasingly penalized. This model recovers classic liquid-democracy behavior in the appropriate limit, satisfies conservation and self-selection properties, and, in contrast to the classical model, the resulting delegation game has pure strategy Nash equilibria when the chain-length penalty is imposed (Bersetche, 2022).

4. Agentic AI implementations

In multi-agent AI, contingent delegation is explicitly a runtime control problem. The “Intelligent AI Delegation” framework operationalizes it through five pillars—Dynamic Assessment, Adaptive Execution, Structural Transparency, Scalable Market Coordination, and Systemic Resilience. Its lifecycle begins with dynamic assessment, proceeds through contract-first task decomposition and market matching, installs a monitoring regime, executes with adaptive coordination, and ends in verifiable completion. External triggers include changes to task specification or cancellation, resource outages or cost spikes, preemption by higher-priority tasks, and detected malicious intent or behavior; internal triggers include performance degradation below SLOs, over-budget resource consumption, failed intermediate verification, and unresponsiveness (Tomašev et al., 12 Feb 2026).

Safe Bilevel Delegation adds a formal runtime safety mechanism to this picture. Its three stated theoretical results are Safety Monotonicity, Inner Policy Convergence, and an Accountability Propagation bound. Safety Monotonicity states that a higher outer safety weight produces a weakly safer inner policy; projected gradient descent on the inner problem converges linearly under standard smoothness assumptions; and accountability along a multi-hop delegation chain is bounded through weights

$G=(V,A)$1

with $G=(V,A)$2 (Sun, 30 Apr 2026).

Contingent delegation in agentic systems also has a protocol layer. Human Delegation Provenance treats it as explicit, verifiable conditional authority: time windows, session binding, scope limits, tool whitelists, network and persistence permissions, and chain-length bounds are attached to a human authorization event and carried in an append-only signed token. Verification is offline and depends only on the issuer’s Ed25519 public key and the current session identifier; the protocol guarantees authenticity and integrity of the record, while semantic enforcement remains local to the agents interpreting scope and header fields (Dalugoda, 6 Apr 2026).

A related governance layer appears in “Delegation Rights,” where the contested object is the mode of account execution rather than platform ownership or data portability. “Certified Delegation” protects delegation only when explicit authorization, revocability, auditability, rate-limit compliance, data minimization, and risk mitigation are verifiably satisfied. The formal mechanism is a threshold rule on risk:

$G=(V,A)$3

so certification acts as a conditional allocation of residual control rather than as a purely technical filter (Zhang et al., 30 Jun 2026).

Privacy-preserving and failure-resilient variants have now been proposed for liquid-democratic settings as well. A sealed delegation regime based on decentralized timed-release encryption hides delegation choices before reveal, while ranked multi-delegation and personal fallback ballots resolve contingencies after reveal time: if the first delegate abstains or fails, the vote automatically shifts to the next-ranked delegate; if all ranked delegates fail or cycles are irresolvable, the personal fallback ballot is counted (Li et al., 2 Jul 2026).

5. Delegation in economics and mechanism design

Economic delegation theory treats contingent delegation as a mechanism by which a principal limits an agent’s discretion while taking the agent’s information or incentives as given. In delegated search, the key result is that structurally simple threshold mechanisms attain sharp approximation guarantees relative to the principal’s own search benchmark. Without independence, there exists a half-infinite interval in $G=(V,A)$4-space that guarantees at least $G=(V,A)$5; under independent atomless marginals for $G=(V,A)$6, an $G=(V,A)$7-threshold yields a $G=(V,A)$8 guarantee; and with full information about $G=(V,A)$9, a decreasing threshold function $S(G)$0 achieves $S(G)$1 with $S(G)$2 (Kleinberg et al., 2018).

The persuasion–delegation equivalence generalizes this insight. Balanced delegation and monotone persuasion are strategically equivalent under the paper’s mapping between marginal utilities, so a contingent delegation policy can be solved by solving the corresponding persuasion problem and then implementing the induced action as a posterior-indexed action set $S(G)$3. In one-dimensional single-crossing environments, this reduces many contingent delegation questions to concavification and monotone partitions over posteriors (Kolotilin et al., 2019).

In veto bargaining, contingent delegation is embodied in the delegation set itself. Full delegation $S(G)$4 is optimal when $S(G)$5 is increasing on $S(G)$6; more generally, interval delegation $S(G)$7 is characterized by boundary and monotonicity conditions together with the first-order condition

$S(G)$8

The substantive implication is unusual for delegation theory: under reasonable conditions, full delegation can nullify the proposer’s bargaining power and still yield ex-post efficiency (Kartik et al., 2020).

Recursive forms of contingent delegation can also be modeled as quitting games and absorbing stochastic games. In that formulation, an agent holding the task chooses among acting, delegating onward, or quitting, and the value of delegating to another agent is the continuation value of that agent’s own optimal future policy. This makes delegation explicitly contingent on the downstream delegate’s anticipated delegation, stopping, or execution behavior, rather than on a one-shot routing decision (Afanador et al., 2018).

6. Effects, controversies, and open problems

A recurrent misconception is that contingent delegation is uniformly beneficial. The literature is more conditional. In secure liquid democracy, delegation improves representational accuracy only when abstention is large and systematically unrepresentative; empirically this is summarized as a recoverable-gap law, with representative-style delegation safer than delegating to a competence elite. The primary benefit of sealed formation is structural rather than directly epistemic: it sharply reduces power concentration, while ranked multi-delegation with personal fallback ballots sharply reduces vote loss under delegate failures (Li et al., 2 Jul 2026).

The epistemic-governance literature is even more skeptical of transfer delegation. It treats “contingent delegation” not as a separate primitive but as a state-dependent orchestration of partial abstention and transfer delegation under competence and dependence tests. On that account, partial abstention can implement optimal log-odds weighting with low coordination burden, while transfer delegation has inherent epistemic weaknesses: it can discard independent information, amplify correlated errors, overweight gurus, and become sensitive to topology and knowledge requirements. Multi-step transfer delegation can reach optimality only under strong assumptions such as connected local competence graphs and divisible votes (Strnad, 7 May 2025).

At the same time, other liquid-democracy work shows that contingent or competence-based delegation need not produce catastrophic concentration. In random transitive delegation models, sufficiently bounded maximum delegated weight together with an average competence uplift is enough for probabilistic “Do No Harm” and “Positive Gain” guarantees. Upward delegation and confidence-based delegation can achieve sublinear or logarithmic concentration bounds with high probability, thereby defending fluid democracy against earlier impossibility results that depended on local-mechanism assumptions (Berinsky et al., 2021).

The limitations are consequently heterogeneous. Some are computational: plurality possible-winner and election-bribery problems are NP-hard or W[1]-hard even on highly restricted DAGs (Bentert et al., 2022). Some are axiomatic: no ranked-path sequence rule is both confluent and copy-robust (Brill et al., 2021). Some are cryptographic or infrastructural: ZK/MPC-based monitoring adds computational overhead, while scalable verification, reputation robustness, transitive liability, and human oversight ergonomics remain open in agentic AI systems (Tomašev et al., 12 Feb 2026). Others are institutional: certification can reduce deadweight loss and restore investment incentives, but the optimal standard is environment-dependent and noisy certification raises questions about suspension, appeal, and capture (Zhang et al., 30 Jun 2026).

Taken together, these literatures portray contingent delegation not as a single doctrine but as a general design pattern for conditional transfer. It can be realized as ranked fallback routing, load-aware balancing, runtime authority attenuation, posterior-indexed action constraints, certified execution rights, or recursive act–delegate–quit policies. This suggests that the enduring research problem is not whether delegation should be contingent, but which contingencies should govern it, how they should be verified, and what trade-offs they induce among autonomy, safety, epistemic quality, accountability, and computational tractability.