Carefully chosen portfolios of AI workflows beat single-workflow deployments: by running multiple complementary workflows and selecting after observing outputs, firms can raise decision accuracy while balancing compute cost, with provable approximation guarantees and dataset gains (e.g., up to +24 percentage points from dual-guided generation on a benchmark).

Main Finding

Deploying multiple agentic-AI workflows as a portfolio (running several workflows in parallel and selecting a final answer after observing their outputs) can improve deployed accuracy versus using the single best workflow — but only when the selector is strong enough to leverage additional correct candidates and when the recurring compute cost of extra executions is justified. The paper introduces an odds-lift index that limits how much selector-aware portfolio design can improve over the best standalone workflow, provides exact and approximate optimization tools for finite and implicit workflow classes, and demonstrates empirical gains on three datasets (ABCD, Schema-Guided Dialogue, HotpotQA).

Key Points

• Problem setting

  • A firm chooses the number of workflow executions (run size) and allocates those executions across workflow types. Each execution produces a candidate answer; a selector then chooses the final output after observing all candidates.
  • Trade-offs: extra executions can (a) uncover correct answers missed by the best workflow, and (b) increase compute cost and add wrong but plausible distractors that worsen selection.

• Selector strength and odds-lift index

  • Introduces an odds-lift index measuring how much the selector increases the odds of returning a correct answer versus random choice among candidates.
  • A finite bound on odds-lift implies a common concave envelope that upper-bounds any portfolio’s selector-aware gains; yields sharp limits on the value of workflow variety and screening rules for when single-workflow deployment is optimal.

• Finite workflow pools

  • For fixed run size the allocation problem is a concave coverage problem.
  • Provides: exact integer programming formulations, linear programming (LP) relaxations, randomized rounding schemes, and computable optimality certificates.
  • Introduces a sparse/grid method to search run size values efficiently while retaining approximation guarantees.

• Implicit (large) workflow classes

  • When feasible workflows cannot be enumerated, the LP relaxation’s dual has only finitely many task-price variables plus one constraint per workflow.
  • Uses a pricing (separation) oracle plus the ellipsoid method to find near-optimal solutions to the relaxation with polynomially many oracle calls (under a weak regularity condition).
  • This yields practical dual-guided workflow generation: the oracle proposes new workflows that violate dual constraints (i.e., high weighted net accuracy considering recurring cost).

• Empirical results

  • Datasets: ABCD (service), Schema-Guided Dialogue (service), HotpotQA (QA).
  • Relative improvements over the best standalone workflow (held-out selector accuracy):
  • ABCD: +3.1 percentage points (best singleton → portfolio), further +3.5 points with dual-guided workflow generation.
  • Schema-Guided Dialogue: +7.5 points (no additional gain from dual-guided generation).
  • HotpotQA: +0.9 points from portfolio; dual-guided generation adds +24.1 points.
  • Selector strength calibrated empirically: substantially above random for the two service domains, positive but weaker on HotpotQA.
  • The number of generated workflows kept varied by domain, consistent with the principle that a workflow is valuable only if its incremental selection benefit outweighs recurring cost.

Data & Methods

• Model setup

  • Workflows: unit types that can be executed multiple times; each execution yields a candidate answer (correct or incorrect).
  • Selector summarized by a recovery curve mapping the number of correct candidates in the pool to probability the selector picks a correct final answer; odds-lift index summarizes selector quality.
  • Objective: maximize expected deployed accuracy minus recurring compute cost (per execution).

• Optimization methods (finite pool)

  • Integer program for joint choice of run size and per-workflow execution counts.
  • LP relaxation to get bounds; randomized rounding converts fractional LP solutions to integer allocations with performance guarantees.
  • Sparse-grid search over run sizes reduces computation for jointly choosing run size with provable approximations.

• Optimization methods (implicit class)

  • Formulate LP with one variable per workflow (infeasible to enumerate).
  • Derive finite-dimensional dual (task-price variables); constraints correspond to workflows.
  • Use a pricing (separation) oracle to find violated constraints (i.e., candidate workflows with positive reduced cost).
  • Apply ellipsoid method with oracle calls to compute ε-approximate solutions in polynomially many steps (polynomial in problem encoding and log(1/ε)).

• Empirical evaluation

  • Calibrate selector recovery curves and odds-lift on ABCD, Schema-Guided Dialogue, HotpotQA.
  • Run portfolio optimization with initial pools; iteratively apply the dual-guided workflow generator (oracle) to propose candidate workflows for inclusion.
  • Measure held-out selector accuracy and report gains versus best single workflow; assess how many workflows are adopted and their marginal contributions.

Implications for AI Economics

• Marginal compute allocation should be decisioned jointly with selection capability

  • The value of an extra workflow execution is not just its marginal probability of being correct, but its marginal impact on the selector’s ability to pick a correct candidate net of recurring compute costs. Firms should weigh compute price-per-execution against the selector’s odds-lift.

• Investment tradeoffs: generating workflows vs improving selector

  • Improving selector quality (increasing odds-lift) can expand the marginal value of portfolio diversity; conversely, stronger selectors reduce the compute budget wasted on distractors. Firms can optimize returns by choosing whether to invest in better selectors, more diverse workflows, or cheaper compute resources.

• Governance and certification

  • Dual certificates and LP bounds provide computable guarantees on portfolio quality relative to the implicit workflow space. These deliverable certificates can be used for procurement decisions, auditing, and governance of deployed agentic systems.

• Product and operational design

  • For many enterprise tasks, maintaining multiple diverse workflows is valuable, but only up to the point where selector confusion or recurring cost outweighs coverage gains. Service providers and SaaS vendors should:
  • Measure selector recovery behavior on representative tasks,
  • Use portfolio optimization to set run size and workflow mix,
  • Use dual-guided generation to direct engineering effort toward workflows that add weighted coverage net of recurring cost.

• Market and competition effects

  • The framework implies diminishing returns to adding many workflows under limited selector lift — suggesting market demand for stronger selectors (meta-models, rankers, verifiers) and for workflow generators that produce highly complementary, low-cost candidates.
  • Pricing of API calls and compute will directly affect optimal portfolio sizes; cheaper compute favors larger portfolios, but only if selectors can exploit extra candidates.

Overall, the paper formalizes the economic trade-offs firms face when deploying agentic AI as portfolios, provides optimization and certification tools for both finite and very large workflow spaces, and quantifies when and how workflow diversity yields measurable deployed-value gains.