Essam Mansour

Research

Enterprise-aware agents: moving agents from individual intelligence to organizational intelligence.

The barrier to enterprise agentic deployment is not model capability. Our SIGMOD 2027 study established this empirically across nine production agent frameworks at fixed LLM and task: orchestration alone changes latency by over 60×, schema-constrained planning costs up to 32 accuracy points through parsing failures rather than reasoning failures, and communication topology swings coordination success from above 90% to below 30%. Not one framework supports controlled knowledge revision. Architecture, not only model capability, governs agent performance. That is a data and systems problem, and we attack it at the data layer across five coupled thrusts.

The active research platform is GEMS (Governed Enterprise Metadata for Agentic Systems), funded through 2030 by NSERC Discovery and NSERC CREATE and released open source through the CoDS-GCS organization. It is staffed by five PhD students and one MSc student recruited specifically against these thrusts. Every system replaces LLM self-judgment with checks against the actual data. See the systems and benchmarks and the people behind each thrust.

THRUST 1

Governed memory and tacit knowledge

Long-term agent memory is a data-management workload, not storage. Every framework we measured treats memory as append-only, so an agent cannot revise what it believes. We treat memory as a governed state machine with typed dependencies, semantic revision, declarative forgetting and multi-tenant privacy, where correctness is a property of the state trajectory rather than of individual records.

Team: Loubna Guelmani (PhD); Dr. Abdelghny Orogat (postdoc)

Related systems: MemState

THRUST 2

Action contracts and governed mediation

Governance must be structural. Agents reach enterprise data only through mediated contracts, so policy is enforced by construction rather than by configuration. Every read, write and promotion of a claim passes a declared gate, which addresses belief laundering: the promotion of agent-generated claims to observed status without a governed gate.

Team: Farbod Aminnezhad (MSc)

THRUST 3

Validation of agent-generated code

Validation must be grounded in data. Assertion-based and derivation-aware checks against the actual data replace probabilistic LLM self-judgment. Classical functional dependencies break under non-deterministic producers, so derivation-aware constraints replace them. CatDB, ReSequel and OCR-APT already follow this pattern.

Team: Qizhi Wang (PhD)

Related systems: CatDB, ReSequel

THRUST 4

Agent security and threat detection

Agent memory is an attack surface. We study multi-agent attack surfaces, memory poisoning and prompt injection under enterprise deployment, building on our work on advanced persistent threat hunting with graph learning and LLMs.

Team: Yomna Gharib (PhD); Ahmed Aly (PhD, co-supervised with Amr Youssef)

Related systems: OCR-APT, MEGR-APT

THRUST 5

Graph foundation models for agentic reasoning

Motif-centric inference over large enterprise knowledge graphs gives agents structured recall at scale. This thrust builds on our GNN-enabled knowledge graph engine, KGNet, and on LLM-guided query-aware inference.

Team: Waleed Afandi (PhD, co-supervised with Ashraf Aboulnaga)

Related systems: KGNet, KG-WISE

Industry and research collaborations

Earlier research

The foundations of the current program, from federated linked data to knowledge graph platforms for data science, are collected under foundational work.