Agentic Causal Bayesian Optimization (A-CBO)

• A-CBO is a framework for sequential, causally informed Bayesian intervention that integrates causal discovery with target optimization.
• It distinguishes itself from standard BO by leveraging structural causal models to guide interventions and manage uncertainty over causal graphs.
• Its practical applications span fields like medicine, ecology, and manufacturing, showcasing the potential of active experimentation in complex systems.

Searching arXiv for the cited Causal Bayesian Optimization papers to ground the article in published work. arxiv_search(query="Causal Bayesian Optimization active causal discovery Gaussian process networks (Kügelgen et al., 2019)", max_results=5) arxiv_search(query="Causal Bayesian Optimization unknown graphs MCBO GACBO constrained causal Bayesian optimization", max_results=10) Across the causal Bayesian optimization literature, a recurring pattern is an active loop in which an agent maintains a causal model, selects interventions, observes outcomes, updates beliefs, and repeats. In that sense, Agentic Causal Bayesian Optimization (A-CBO) denotes sequential, Bayesian, causally informed intervention selection for either causal discovery or target optimization. The underlying object is typically a structural causal model (SCM), the action space consists of interventions such as $do(X_j=x)$ or $do(X_s=x_s)$, and the decision rule is an acquisition or utility criterion defined over interventional consequences rather than over a purely black-box input space (Kügelgen et al., 2019, Sussex et al., 2022, Mukherjee et al., 2024).

1. Conceptual scope and defining characteristics

A-CBO departs from standard Bayesian optimization in three linked ways. First, interventions are interpreted through an SCM or causal graph, so actions are not generic design points but manipulations of variables embedded in directed dependencies. Second, uncertainty is not restricted to a reward surface; it may concern causal graph structure, nodewise mechanisms, direct parents of the target, or shared causal parameters across intervention families. Third, the sequential objective may be causal discovery, reward optimization, or an explicit hybrid in which structure learning is pursued only insofar as it improves future decisions (Kügelgen et al., 2019, Durand et al., 25 Mar 2025, Mukherjee et al., 2024).

This framing also distinguishes A-CBO from passive causal inference. Observational data may initialize priors or identify parts of the model, but the central mechanism is active experimentation. The system chooses what intervention to perform next using its current posterior uncertainty, not a fixed experimental schedule. In the discovery setting, the action is selected to reduce uncertainty about the causal graph; in the optimization setting, it is selected to improve a downstream target variable; in several recent variants, both motives are present simultaneously (Kügelgen et al., 2019, Wang et al., 2024, Sussex et al., 2022).

A common misconception is that causal Bayesian optimization is simply standard BO with a causal prior. The literature is more diverse. Some methods optimize expected posterior entropy reduction over graphs, some optimize Bayes-factor-based decisiveness, some optimize interventional reward under a known graph, and some explicitly treat graph discovery as a subtask of optimization rather than as a separate preliminary stage (Kügelgen et al., 2019, Wang et al., 2024, Mukherjee et al., 2024).

2. Emergence from active causal discovery

An early and explicit precursor of A-CBO appears in targeted causal discovery with continuous nonlinear variables and Gaussian process networks. In that formulation, the environment is an SCM over real-valued variables $\mathbf X=\{X_1,\dots,X_d\}$ with unknown causal DAG $G^*$ and additive noise mechanisms

$X_i = f_i(\mathbf{Pa}_i^{G^*}) + \epsilon_i.$

The agent starts from few observational measurements, then performs targeted experiments of the form $do(X_j=x)$ and observes samples from $P(\mathbf{X}_{-j}\mid do(X_j=x))$ (Kügelgen et al., 2019).

The discovery objective is Bayesian. A posterior over graphs and parameters is maintained, and the next experiment is chosen to maximize expected information gain about the graph. The core criterion selects the target variable and intervention value by

$$(j^*,x^*) = \arg\max_{j,x} \sum_{G\in\mathcal G} P(G)\int P(\mathbf x_{-j}\mid G,do(X_j=x)) \log P(G\mid \mathbf x_{-j},do(X_j=x)) \,d\mathbf x_{-j}.$$

This formulation matters because observational data often identify only a Markov equivalence class, whereas interventions alter the factorization and can distinguish among equivalent DAGs (Kügelgen et al., 2019).

The difficulty is twofold: the expected information gain integral is hard to compute, and the intervention value $x$ lies in a continuous, uncountable domain. The proposed solution models structural mechanisms with Gaussian process priors, making graph evidence and predictive posteriors tractable enough for Monte Carlo estimation, and then uses Bayesian optimization with a GP-UCB acquisition rule to optimize the intervention value over $\mathcal X_j$ (Kügelgen et al., 2019). The resulting loop is recognizably agentic: observational initialization, posterior computation, pre-experimental simulation of interventions, selection of the most informative experiment, posterior update, and repetition.

The same general discovery agenda has also been reformulated around hypothesis testing rather than entropy reduction. "Bayesian Intervention Optimization for Causal Discovery" proposes a Bayes-factor-inspired criterion, the Probability of Decisive and Correct evidence (PDC), and selects the intervention value $do(X_s=x_s)$0 to maximize the expected probability that the next experiment will produce decisive and correct evidence for the true causal hypothesis (Wang et al., 2024). This directly challenges the assumption that information gain is always the right objective for active causal discovery. In this line of work, the active loop remains the same, but the utility is tied to statistical decisiveness rather than posterior entropy.

3. Core mathematical architecture

Despite their diversity, A-CBO methods share a common architecture. A causal model, explicit or implicit, defines interventional queries; a Bayesian posterior quantifies uncertainty; an acquisition or utility criterion ranks candidate interventions; an action is executed; new interventional data are incorporated; and the posterior becomes the prior for the next round. What changes across methods is the object of uncertainty and the decision criterion.

For discovery-oriented methods, the utility is often expected information gain or a Bayes-factor-derived quantity. For optimization-oriented methods, the central object is an interventional response such as

$do(X_s=x_s)$1

and the optimization problem is to choose $do(X_s=x_s)$2 maximizing or minimizing this response over a family of candidate intervention sets (Javidian, 31 May 2026, Durand et al., 25 Mar 2025).

Surrogate construction is equally heterogeneous. Gaussian process networks model nonlinear continuous mechanisms nodewise and induce graph posteriors through GP marginal likelihoods (Kügelgen et al., 2019). Model-based CBO learns the full SCM by placing a GP prior on every node mechanism and propagating epistemic uncertainty through the graph (Sussex et al., 2022). Unknown-graph variants maintain a posterior over parent sets of the target and use law-of-total-expectation and law-of-total-variance to integrate structure uncertainty into the surrogate mean and variance (Durand et al., 25 Mar 2025). Graph-coupled causal BO instead parameterizes all intervention responses by a shared identifiable causal parameter vector $do(X_s=x_s)$3 and induces cross-intervention covariance via

$do(X_s=x_s)$4

where $do(X_s=x_s)$5 is the Jacobian with respect to $do(X_s=x_s)$6 and $do(X_s=x_s)$7 is the posterior covariance of the shared parameters (Javidian, 31 May 2026).

This suggests that A-CBO is better understood as a family of Bayesian decision procedures over interventional queries than as a single algorithm. What unifies the family is the closed-loop coupling of causal modeling and action selection.

4. Known-graph optimization regimes

When the causal DAG is known, the central technical question becomes how much of the causal system should be modeled and how the graph should restrict the intervention space. Three representative designs are prominent.

Method	Graph assumption	Characteristic mechanism
MCBO	Known DAG	Learns full SCM and acts optimistically
cCBO	Known DAG	Prunes intervention sets and enforces constraints
GC-CBO	Known DAG	Transfers information across interventions via shared causal parameters

Model-based Causal Bayesian Optimization (MCBO) assumes a known DAG with unknown node mechanisms and noisy observations. Rather than modeling only intervention-reward pairs, it learns a probabilistic model for every mechanism in the SCM and selects actions by optimism over plausible causal systems:

$do(X_s=x_s)$8

Its cumulative-regret analysis yields the first non-asymptotic bounds for CBO, with regret scaling in terms of graph depth $do(X_s=x_s)$9, maximum indegree $\mathbf X=\{X_1,\dots,X_d\}$0, the number of nodes $\mathbf X=\{X_1,\dots,X_d\}$1, confidence width $\mathbf X=\{X_1,\dots,X_d\}$2, Lipschitz constants, and GP information gain $\mathbf X=\{X_1,\dots,X_d\}$3 (Sussex et al., 2022). A plausible implication is that known causal structure can replace an otherwise high-dimensional black-box reward model by a compositional world model whose complexity depends on graph structure.

Constrained Causal Bayesian Optimization (cCBO) studies known-graph intervention optimization when the target must be optimized subject to inequality constraints on other variables. Its first stage, cMISReduce, discards intervention sets that are provably redundant using ancestor reasoning in the mutilated graph and observational-data-based reducibility and null-feasibility tests. Its second stage fits GP surrogates for target and constraint effects and uses constrained expected improvement, normalized by intervention cost, to choose the next action (Aglietti et al., 2023). Here the agentic element is algorithmic rather than rhetorical: the method first reasons symbolically about admissible intervention sets, then performs adaptive BO on the reduced action space.

Graph-coupled Causal Bayesian Optimization (GC-CBO) addresses a different inefficiency in known-graph CBO: standard approaches often learn each intervention response almost in isolation. GC-CBO couples intervention responses through a shared causal parameterization and obtains a low-rank causal kernel whose rank is bounded by $\mathbf X=\{X_1,\dots,X_d\}$4 rather than by the size of the intervention menu. The resulting information-gain bound is $\mathbf X=\{X_1,\dots,X_d\}$5, and the regret bound decomposes into optimization error, causal estimation error, and intervention-set selection error (Javidian, 31 May 2026). The strongest gains are reported when direct interventions on the target’s parents are unavailable and sparse interventional data must be reused across a large family of candidate interventions.

5. Unknown and partially known graph settings

A-CBO does not require a fully known graph in all formulations. One major line of work argues that, for optimization, learning the entire DAG may be unnecessary. Causal Bayesian Optimization with Unknown Graphs (CBO-U) restricts attention to hard interventions and to settings in which $\mathbf X=\{X_1,\dots,X_d\}$6 has no causal effect on features, the direct parents of $\mathbf X=\{X_1,\dots,X_d\}$7 are intervenable, and $\mathbf X=\{X_1,\dots,X_d\}$8 is not influenced by confounders. Under these assumptions, direct parents of the target are sufficient for optimization, and the method maintains a Bayesian posterior over candidate parent sets rather than over full graphs (Durand et al., 25 Mar 2025). In the linear case it derives a closed-form posterior update for parent sets; in the nonlinear case it approximates a GP target mechanism with random Fourier features so that Bayesian linear regression machinery remains available. The method alternates between parent-set inference, surrogate construction under parent uncertainty, intervention selection, and posterior updating, and it is reported to scale to graphs of at least 100 nodes (Durand et al., 25 Mar 2025).

A broader alternative is Graph Agnostic Causal Bayesian Optimisation (GACBO), which studies cumulative-regret CBO when the graph is unknown or partially known, both for hard interventions in SCMs and for soft interventions in function networks. GACBO maintains posterior uncertainty over graph structure and node functions, scores graphs with Bayesian scoring, forms plausible graph sets, and uses a UCB-style rule that is optimistic over both plausible graphs and plausible functions (Mukherjee et al., 2024). The paper explicitly frames causal discovery as a subtask of optimization: structural distinctions are worth learning only when they can change the best intervention or improve cumulative reward. In simulated and real-world experiments, GACBO is reported to approach near-oracle known-graph performance within about 100 rounds, although finite-sample regret guarantees for the full graph-agnostic setting are left to future work (Mukherjee et al., 2024).

These two directions correct another common misconception: unknown-graph A-CBO need not mean full causal discovery. One strategy learns only the local neighborhood of the target; another learns graph structure adaptively and only to the extent required for reward improvement.

6. Extensions, applications, and limitations

Several extensions broaden the operational scope of A-CBO. Adversarial Causal Bayesian Optimization (ACBO) generalizes CBO to environments in which external agents or exogenous events also intervene on the system. Its algorithm, CBO-MW, combines optimistic counterfactual reward estimation through a known causal graph with multiplicative weights, and obtains bounded regret against an adversarial sequence of external actions (Sussex et al., 2023). This moves A-CBO from stationary intervention design toward robust online decision-making under non-stationarity.

Multi-Source Causal Bayesian Optimization (MSCBO) combines causal intervention-set pruning with multi-source BO. The agent chooses not only what to intervene on, but also which information source to query, and whether to observe or intervene, using a cost-sensitive knowledge gradient and an $\mathbf X=\{X_1,\dots,X_d\}$9-greedy observation-versus-intervention rule. In benchmarks including PSA, E. coli, crop yield, honey yield, and a multi-armed bandit graph, the reported advantage is strongest when causal pruning via POMIS substantially reduces the intervention dimension and source selection lowers operational cost (Jacobs et al., 16 Feb 2026).

Across the literature, application domains include medicine, ecology, and manufacturing, as well as healthcare PSA optimization, epidemiology for HIV viral load minimisation, protein-signaling and E. coli networks, and shared mobility demand adaptation (Sussex et al., 2022, Aglietti et al., 2023, Mukherjee et al., 2024, Sussex et al., 2023). The practical motivation is consistent: experiments are expensive, interventions are structured by causal mechanisms, and computational planning is cheaper than physical experimentation.

The main limitations are equally consistent. Many methods still assume the true DAG is known, and graph misspecification can invalidate search-space reduction or prior construction (Aglietti et al., 2023). CBO-U is explicitly restricted to hard interventions and to target variables with no outgoing effects and no confounding (Durand et al., 25 Mar 2025). GC-CBO’s strongest theory is for identifiable linear Gaussian models; in nonlinear settings it replaces exact low-rank covariance with a path-Lipschitz propagation bound (Javidian, 31 May 2026). GACBO studies unknown graphs under cumulative regret but does not yet provide finite-sample regret guarantees for the full graph-agnostic setting (Mukherjee et al., 2024). MCBO provides non-asymptotic regret analysis, but practical acquisition optimization is nonconvex and computationally heavy (Sussex et al., 2022).

Taken together, these results portray A-CBO not as a single standardized algorithm but as a research program centered on causally structured, Bayesian, sequential intervention selection. The unifying principle is that an optimizer should not learn intervention effects as unrelated black-box tasks. It should maintain a causal world model, propagate epistemic uncertainty through that model, and choose interventions according to the decision criterion that is most aligned with the task at hand: entropy reduction for graph learning, decisive Bayes-factor evidence for hypothesis testing, reward improvement under cumulative regret, feasibility-constrained utility, transfer across intervention families, or robustness to external agents (Kügelgen et al., 2019, Wang et al., 2024, Javidian, 31 May 2026).