Delegation Over Experiments

• Delegation over experiments is a framework where decision makers transfer control over experimentation to agents while embedding constraints, oversight, and verification mechanisms.
• The analysis encompasses formal models like balanced delegation and delegated search, which reveal trade-offs between restricted decision sets and controlled information disclosure.
• Applications span AI systems to organizational testing, with empirical evidence underscoring the need for explicit governance, traceability, and calibrated recovery protocols.

Delegation over experiments denotes a class of problems in which a principal, designer, regulator, or coordinating system does not directly choose an experiment, execute a task, or evaluate every output itself, but instead constrains, routes, or supervises another actor that does. In recent work, the delegated object may be a Blackwell experiment, a sequential trial design, a recursive task-routing policy, a multi-agent workflow, or end-to-end ownership of online controlled experiments. Across these settings, the common structure is transfer of authority across an incentive or trust boundary, with quality, disclosure, verification, attestation, and recovery becoming central design variables rather than peripheral implementation details (Bilotta et al., 11 Mar 2026, Adusumilli et al., 15 May 2026, Prakash, 15 Mar 2026, Kaufman et al., 2017).

1. Formal foundations and canonical models

At the most abstract level, delegation over experiments is a principal–agent design problem over either action menus or experiment menus. Kolotilin and Zapechelnyuk show that balanced delegation and monotone persuasion are strategically equivalent in a one-dimensional environment: restricting the agent’s decision set and restricting the agent’s information by a monotone partition are two representations of the same optimization problem, linked by payoff transformations that swap the roles of state and decision through marginal utilities (Kolotilin et al., 2019). This equivalence is important because it turns many “who may act?” questions into “what may be learned?” questions, and conversely.

Kleinberg and Kleinberg’s delegated-search model makes the same point in a different language. There the agent conducts a hidden search over candidate solutions, which is naturally interpretable as a hidden experimental campaign. The principal cannot observe the full search path and can only commit to an eligible set of acceptable proposals. Even under this severe informational restriction, simple threshold mechanisms obtain strong approximation guarantees relative to first-best search: a universal $1/2$ guarantee, $1-1/e$ under independent atomless utilities, and approximately $0.745$ when the agent’s utility $y$ is observable (Kleinberg et al., 2018). In this formulation, an experiment is valuable not because it is intrinsically informative, but because it induces a proposal that survives an ex ante acceptance rule.

Recursive delegation further enlarges the state space. In the quitting-games formulation, the value of delegating to $j$ is not a local estimate of $j$’s competence but a continuation value,

$$V_i=\max\Bigl\{U_i^{\mathrm{exec}},\ \max_{j\in \mathcal N(i)}\bigl(-c_{ij}+\mathbb E[V_j]\bigr)\Bigr\},$$

because the delegate may delegate again (Afanador et al., 2018). The reported simulations indicate that quitting-game-based policies outperform policies that do not explicitly account for the recursive nature of delegation (Afanador et al., 2018). Delegation over experiments is therefore not exhausted by one-shot assignment; it often concerns the control of downstream experimental opportunities generated by prior delegation choices.

2. Delegated information design and strategic experimentation

In the most direct use of the phrase, “Delegated Information Provision” studies a designer who cannot produce information directly and instead chooses a restriction $F\in MPC(H)$, while the experimenter may choose any garbling in $MPC(F)$ (Bilotta et al., 11 Mar 2026). Incentive compatibility is self-enforcement: an experiment is incentive compatible if and only if $F\in BR(F)$ (Bilotta et al., 11 Mar 2026). Using Dworczak and Martini’s price-theoretic characterization, the paper shows that this is equivalent to the existence of a convex, continuous function $1-1/e$0 such that the support of $1-1/e$1 lies in the contact set where $1-1/e$2 (Bilotta et al., 11 Mar 2026). Under $1-1/e$3-shaped experimenter preferences, the maximally incentive-compatible experiments are exactly double censorship: low states are revealed, an intermediate interval $1-1/e$4 is pooled to $1-1/e$5, and the upper tail $1-1/e$6 is pooled to $1-1/e$7, with $1-1/e$8 and $1-1/e$9 (Bilotta et al., 11 Mar 2026). The substantive conclusion is that full delegation is not optimal; a nontrivial restriction strictly benefits the designer because shrinking the upper pooling region yields a first-order gain, whereas the informational loss from introducing a small intermediate pooling region is second-order (Bilotta et al., 11 Mar 2026).

A different delegated-experiment problem appears in the regulator–experimenter model of persuasive trial design. Bob chooses a posterior distribution $0.745$0 subject to Bayes plausibility,

$0.745$1

and a regulator-imposed welfare floor,

$0.745$2

without the regulator needing to know Bob’s private payoffs or experimentation cost (Adusumilli et al., 15 May 2026). In the sequential Gaussian specialization, the striking result is that the optimal sampling rule is always Neyman allocation,

$0.745$3

independent of objectives; delegated distortion enters only through the stopping rule, via the additive approval payoff $0.745$4 and asymmetric stopping boundaries satisfying $0.745$5 (Adusumilli et al., 15 May 2026). In the numerical calibration, the welfare-threshold mechanism matches the welfare of a standard RCT with $0.745$6, corresponding to about 48% fewer observations on average (Adusumilli et al., 15 May 2026). Delegation is therefore constrained not by forbidding adaptive design, but by limiting persuasive stopping behavior.

Private sequential experimentation introduces another failure mode: verifiable outcomes without verifiable completeness. Under continuous disclosure, dynamic private experimentation can be reformulated as a static signal-design problem, and equilibrium-feasible posterior distributions are exactly mean-preserving spreads of the prior supported on additional-learning-proof beliefs $0.745$7 (Lou, 2023). Posteriors that would invite profitable hidden refinement and right-truncated disclosure are excluded. In delegated experimentation, this means that feasible information structures are constrained not only by Bayes plausibility, but also by robustness to concealed extra learning (Lou, 2023).

3. Protocolized delegation in AI systems

In multi-agent LLM systems, delegation over experiments becomes a routing problem under untrusted capability signals. “The Provenance Paradox in Multi-Agent LLM Routing: Delegation Contracts and Attested Identity in LDP” shows that when routers monotonically prefer higher self-reported quality, strategic inflation can invert selection pressure and make routing worse than blind selection (Prakash, 15 Mar 2026). In the reported experiments, self-claimed routing underperformed random in simulation, $0.745$8 versus $0.745$9, and in real-model validation, $y$0 versus $y$1; attested routing reached near-optimal performance with $y$2 and $y$3 (Prakash, 15 Mar 2026). Across 36 sensitivity configurations, self-claimed routing was worse than blind in $y$4, especially when dishonest delegates constituted at least $y$5 of the pool and inflation magnitude was medium to high (Prakash, 15 Mar 2026).

The protocol response is to make delegation governable. The LLM Delegate Protocol extension distinguishes self_claimed, runtime_observed, issuer_attested, and externally_benchmarked quality; adds optional delegation contracts with objective, success_criteria, budget, deadline, safety_constraints, max_delegation_depth, and failure_policy; and replaces free-form string failures with typed LdpError categories including runtime, transport, policy, capability, quality, identity, and session (Prakash, 15 Mar 2026). Contract violations specifically yield category: policy, code: CONTRACT_VIOLATED, with partial_output preserved (Prakash, 15 Mar 2026). All new fields are optional, old TASK_SUBMIT messages still deserialize unchanged, and delegates unaware of the new fields can ignore them (Prakash, 15 Mar 2026). The reported overhead is modest: message size rises from 972 bytes to 1,497 bytes when a contract is added, contract validation costs 0.45 $y$6, and serialization adds approximately 4 $y$7, which the paper characterizes as sub-microsecond validation overhead relative to 500 ms–5 s LLM latency (Prakash, 15 Mar 2026).

This protocol literature changes the meaning of “experiment.” The object being delegated is no longer only an information structure in the Blackwell sense; it is a bounded, typed, provenance-aware execution under explicit recovery semantics. Delegation over experiments in such systems therefore includes control of who may act, on what claim basis, within what budget, and with what post-failure behavior.

4. Workflow architecture as experimental treatment

Recent work treats workflow structure itself as the experimental condition. “Agents for Experiments, Experiments for Agents” introduces SEED, a typed, attributed, directed relational graph $y$8 whose nodes are human actors $y$9 and agentic AI actors $j$0, and whose flows are typed as content $j$1, control $j$2, or iterative interaction $j$3 (Zhang et al., 18 May 2026). Governance is not appended afterward; it is encoded directly on edges through moderators

$j$4

where protocols, incentives, and information design constrain what a delegated relation may do (Zhang et al., 18 May 2026). Structural mutations—adding actors, redirecting edges, converting content flow to control flow, or returning final authority—become the experimental manipulation itself (Zhang et al., 18 May 2026).

The feasibility test in medical triage illustrates the gain from such encoding. A local library of 65 coded records collapsed into 25 condition-graph templates was used to compare graph-blind generation with structure-first generation based on eight citation-backed anchor graphs and five graph operations (Zhang et al., 18 May 2026). Structure-first generation improved all 12 atomic dimensions from 3.656 to 6.275, design-quality average from 3.942 to 6.150, graph-discovery average from 3.280 to 6.475, and reduced post-hoc graph clarification need from 4.900 to 1.000; in a simulated reviewer audit, 36 of 40 SEED candidates were rated “Keep,” versus 0 of 40 graph-blind candidates (Zhang et al., 18 May 2026). The experimental contribution is not a clinical effect estimate but a representation result: delegation becomes comparable because authority, mediation, escalation, and oversight are explicitly encoded rather than described only in prose.

This structural view aligns with recursive delegation models. In quitting-game-based delegation, downstream delegation matters empirically because the value of assigning a task to an intermediary depends on the intermediary’s future delegation policy, not merely on immediate competence (Afanador et al., 2018). A plausible implication is that many “AI versus human” experiments are underspecified when they ignore who may delegate onward, who may veto, and where iterative feedback loops occur.

5. Organizational, behavioral, and collective-choice evidence

Organizational experimentation offers a large-scale instance of delegated experimental authority. Booking.com reports more than a thousand concurrent experiments daily across customer-facing web and mobile products, tools for hoteliers, customer service phone lines, and internal systems, and presents this as democratization and decentralization of experimentation and decision making (Kaufman et al., 2017). End-to-end ownership is distributed: experiment owners formulate hypotheses, implement targeting and placement logic in business code, launch and monitor the test, analyze results, and make the judgment call on what to do next, while the platform centralizes reusable mechanics such as randomization, recording, and reporting (Kaufman et al., 2017). Trust in this delegated regime is maintained by dual independent aggregation pipelines, one with less than five minutes of delay and one with daily batch updates, controlled-input experiments, AA experiments, and safeguards such as hiding comparative statistics when selective attrition is detected (Kaufman et al., 2017). Delegation here is not laissez-faire; it is institutionalized through infrastructure, monitoring, and training.

Human-subject experiments show that delegation changes both objective and subjective outcomes. In a randomized study with 196 participants performing distorted ImageNet classification, AI delegation produced 84.51% accuracy in the delegation condition and 83.73% in hidden delegation, compared with 67.13% in the human-alone condition; combined human–AI team performance was 80.01%, compared with 75.83% for AI alone on the full test set (Hemmer et al., 2023). Task satisfaction also increased, and self-efficacy statistically mediated the treatment effect, partially for performance and fully for satisfaction (Hemmer et al., 2023). This is evidence that delegated task composition is itself a behavioral intervention, not merely an optimization layer over fixed human performance.

Delegation, however, is often overused. In an online financial-choice experiment with 589 participants, pooled delegation rates were about 50.4% in trivial tasks, 51.3% in simple tasks, and 61.6% in complex tasks (Freer et al., 2023). The trivial-task result is particularly informative because delegation there has no instrumental value; the paper interprets this as strong evidence for blame shifting or decision avoidance, while the additional rise in the complex task supports a role for decision costs (Freer et al., 2023). In voting experiments on liquid democracy, delegation underperformed both universal majority voting and abstention; in the lab experiment this was due to overdelegation, and in the perceptual task subjects substantially overestimated the precision of the better informed voters (Mooers et al., 2022). Computational experiments on ranked liquid democracy display the same trade-off in a different idiom: BFD minimizes path length and MaxWeight, Borda and DFD favor better-ranked paths and lower unpopularity, and MinSum lies between these poles (Brill et al., 2021). Across these literatures, the recurring theme is that delegation’s theoretical gains are narrow and sensitive to how agents perceive competence, responsibility, and fallback structure.

6. Observability, oversight, and unresolved tensions

Delegated execution creates an observability problem that standard telemetry does not solve. “Observability for Delegated Execution in Agentic AI Systems” proves that delegation assignment is not identifiable from conventional observables $j$5 when delegation membership is not explicitly encoded: the observation function

$j$6

discards the authority relation, so multiple incompatible authorization assignments induce identical logs and traces (Mishra et al., 8 Jun 2026). The proposed remedy is an agent-aware observability substrate composed of a lightweight gateway and a Common Information Model with fields such as delegation_id, delegation_parent_id, principal.user_id, agent.agent_id, tool.name, action.semantic, and explicit unbound/unknown states (Mishra et al., 8 Jun 2026). This makes delegation-scoped reconstruction a first-class query object rather than a heuristic correlation exercise.

Minimal-oversight theory turns such traceability into a governance calculus. The Minimum Sufficient Oversight Principle minimizes

$j$7

subject to a delivery constraint, yielding the water-filling solution

$j$8

on the Fisher information manifold (Azevedo, 4 Jun 2026). The same framework defines an effective autonomy buffer,

$j$9

and a drift-dominated autonomy time,

$j$0

while identifying masking as a structural pathology through $j$1: corrected performance can conceal the competence signal required to calibrate trust (Azevedo, 4 Jun 2026). For delegated experiments, this implies that post-correction success rates are insufficient governance evidence unless raw and corrected outcomes are logged separately.

At the human-oversight level, “Delegation and Verification Under AI” sharpens the same point. Rational workers choose among exactly three workflows—manual work, pure delegation, and verified delegation—and AI induces phase transitions in which arbitrarily small differences in verification ability lead to sharply different behaviors (Huang et al., 3 Mar 2026). Worker quality is institution-centered,

$j$2

not worker-centered, and the formal result is that AI can amplify workers with strong verification reliability while degrading institutional worker quality for others who rationally over-delegate and reduce oversight, even when baseline task success improves and no behavioral biases are present (Huang et al., 3 Mar 2026). The SEED commentary identifies parallel tensions around novelty, replication, validity, diversity of inquiry, and accountability in AI-assisted experimental design (Zhang et al., 18 May 2026). Taken together, these results suggest that delegation over experiments remains tractable only when provenance, attestation, verification, and traceability are encoded as first-class design variables rather than treated as downstream administrative detail.

Delegation over experiments therefore names not a single technique but a recurring problem of controlled authority transfer in the production of evidence, decisions, and workflow outcomes. Its modern literature converges on a common lesson: better delegation does not arise from delegation alone, but from explicit restrictions on admissible experiments, calibrated routing signals, typed recovery semantics, and governance mechanisms that preserve the difference between what a delegate achieved unaided and what a supervisory system repaired after the fact.