From Implicit to Explicit Causal Models in BDI Agents: A Proof-of-Concept with Causal Discovery and Adaptive Policy

Meltem Gullusac, Baris Tekin Tezel, MOHARRAM CHALLENGER

Keywords: BDI agents, causal reasoning, structural causal models, causal discovery, Jason/AgentSpeak, Thompson Sampling, latent confounders

TL;DR: We introduce CAUSALBDI, a framework that enables BDI agents to transition from implicit correlational beliefs to explicit causal reasoning using the FCI algorithm and Thompson Sampling to mitigate latent confounding.

Abstract: Belief-Desire-Intention (BDI) agents encode causal assumptions implicitly within their plan libraries and belief-update rules, without formally representing or reasoning over causal structure. This limits their ability to detect confounding, distinguish correlation from causation, and adapt to environments with latent variables. We present CausalBDI, an architectural pattern that augments Jason/AgentSpeak agents with an explicit structural causal model (SCM) maintained by an external causal inference server. The agent progresses through an epistemic lifecycle: (1) a naive phase with epsilon-greedy exploration to collect randomised-policy data, (2) causal discovery via FCI on data from random action selection to detect latent confounders, (3) an epistemic transition where beliefs shift from correlational to causal, and (4) a causal policy phase using proxy-stratified estimation, sensitivity analysis, and Thompson Sampling. We evaluate the architecture on a navigation scenario where a proxy variable (yellow warning light) misleads naive agents into suboptimal behaviour: over 30 independent runs, the causally-aware agent reduces accident rate by 25% compared to the naive baseline ($p < 0.001$), closing 42% of the gap to an oracle with direct access to the latent confounder. Our work provides a concrete proof-of-concept for integrating causal learning and reasoning into the BDI reasoning cycle, bridging the gap between agent programming and causal inference.

Paper Type: Full (minimum of 10 pages and a maximum of 16 excluding references)

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 14