Knowledge Graphs Give Enterprise AI Agents Shared Context and Runtime Controls
- Cargo
- AI Tools & Apps
- Shipped
- Shipper
- Victoria Gayton
- Readers
- 497
GraphSummit teams describe how shared knowledge graphs support traceable answers, governed agents, safer code changes and verified financial workflows.

Enterprise AI agents need more than access to business data. They need to understand how that data relates to the organization’s systems, rules and established sources of truth. At GraphSummit 2026, practitioners described knowledge graphs as a shared context layer that can help agents work with that information while making their actions easier to check.
Without shared context, developers often encode business knowledge in prompts for each new agent. That approach repeats work and can leave agents with inconsistent instructions. Jesús Barrasa, field chief technology officer of generative AI at Neo4j, argued that a central knowledge layer could give agents consistent context and make answers more explainable by showing which information and relationships informed them.
That traceability matters as organizations move from experiments to workflows where agents make decisions or take actions. The GraphSummit discussions covered governance, large software changes and financial services, with each example showing a different way to connect agent behavior to knowledge people can inspect.
Governance that runs alongside agents
For Collibra CEO and co-founder Felix Van de Maele, the governance challenge changes when agents act autonomously. Decisions that people once made over longer periods may now happen in seconds, so governance has to operate while the agent is working.
Collibra introduced Live Map, Maestro, Guardian Agents and Agent Contracts, capabilities built on Neo4j’s graph database. Together, they are intended to give organizations ways to direct agents toward appropriate information and keep their behavior within defined boundaries. Maestro can support governance steward agents that automate parts of governance work. Van de Maele said it can automate up to 80% of that work, helping teams keep pace with new AI workloads.
Agent Contracts address behavior directly: they monitor agent actions and can block activity that breaks predefined rules. This kind of runtime control is designed to make agent boundaries enforceable during execution, rather than relying only on instructions supplied before a task begins. The practical question for deployments is how teams define those rules and keep them aligned with changing business processes.
Mapping dependencies before changing code
Large software systems pose a different context problem. A seemingly small code change can affect distant components, while asking an agent to absorb millions of lines of code can overwhelm the task. Blitzy uses knowledge graphs to map code dependencies and help its agents assess the risks of changes across a system.
Blitzy runs thousands of coding agents at once to support large-scale system changes. Director of engineering Neeraj Deshmukh said the graph gives agents a structured view of relevant code and relationships. The workflow also includes action plans approved by people and automated testing.
Blitzy adds agents that monitor other agents’ work against the agreed specification and plan. That monitoring is meant to catch drift or fabricated output before it spreads through a change. The layered process offers a practical pattern for coding systems: map dependencies, set a human-approved plan, then check agent activity and test the result.
Checking financial workflows against trusted sources
At Raymond James Financial, the knowledge graph helps connect agent work to established data sources. Vice president of technology Naveen Mahankali described using agents to build outputs and compare them with the company’s “golden sources.” Data from different systems is brought into Neo4j, where the team runs evaluations against those reference sources before launch. Mahankali said the checks are intended to ensure results are 100% accurate before deployment.
That figure describes the company’s stated verification target for its workflow; it doesn’t establish how the system performs across every use case. The approach nevertheless illustrates why a graph can be useful in an enterprise setting: it gives teams a way to relate data from multiple places and evaluate agent output against known references.
Raymond James is also exploring agents to help financial advisors serve clients more efficiently. The firm has launched Raimond, its proprietary AI assistant, which vice president of AI engineering and data science Jeff Griffith said has seen significant adoption and is expected to improve over time.
Griffith described a shift in how employees work: people may spend less time on routine tasks and more time managing agents that handle them. He also pointed to a change in the skills the company values, from simply completing tasks toward identifying and solving meaningful problems.
Shared context is only part of the system
The examples point to a common design principle. A knowledge graph can give agents structured business context, but organizations still need controls around what agents may do, checks on the outputs they produce and ways to measure whether the results meet their standards. Those supporting mechanisms differ by task: contracts and governance agents for oversight, dependency maps and tests for code, and reference data with evaluations for financial workflows.
The next test for these systems is whether shared context and runtime checks can keep pace as organizations add agents and expand their responsibilities. Graphs can make relationships visible; teams still have to decide which sources are trusted, which actions are allowed and how success is verified.