Knowing Before Doing with Sudhir Hasbe

Summary

Sudhir Hasbe frames agentic AI as a knowledge problem first and an action problem second: systems should know before they do. He argues that both too little context and too much context can mislead an agent, and that memoryless systems cannot improve because they forget the outcomes of prior attempts. The discussion turns to enterprise AI failure modes, especially organizational unreadiness and poorly structured data, rather than model quality alone. Hasbe describes a self-learning pattern in which unresolved common requests are fed back into a knowledge graph as shared terminology, helping the system improve over time. The broader strategic claim is that graphs are about boundaries and useful context, not simply accumulating more data or enabling more use cases.

Notes

• Agentic systems should begin with knowledge acquisition rather than execution.
• Too much knowledge can be as misleading as too little knowledge.
• A memoryless agent cannot improve because it has no durable learning loop.
• Self-learning design can push unresolved common terms back into a graph.
• Enterprise AI projects still fail because organizations are not ready for these systems.
• Data structure, not just model capability, is presented as a core bottleneck.
• The episode treats knowledge graphs as an operational memory layer for agents.
• Agents need a way to retain outcomes from prior interactions.
• The design goal is learning behavior, not a one-shot answer engine.
• Graph updates can be triggered by repeated failures on common requests.
• Repeated unresolved requests become a source of vocabulary expansion.
• Knowledge management is positioned as a prerequisite to dependable autonomy.
• The team emphasizes a narrow, grounded context rather than maximal context windows.
• The knowledge problem is framed as an enterprise readiness problem.
• A system that cannot resolve common terms needs a feedback path into its ontology or graph.
• The episode treats action without context as premature.
• A learning agent should update its vocabulary from user behavior.
• The discussion implies that context curation is a systems engineering problem.
• Persistent memory is required for iterative improvement.
• The graph is used as a shared source of terminology across repeated interactions.
• Enterprise failures are linked to unstructured data conditions.
• The episode argues for systems that can generalize from failures, not merely respond.
• Repeated user language can become machine-readable structure.
• The architecture described is closer to an adaptive knowledge system than a stateless chatbot.
• The value proposition is operational learning over raw model intelligence.
• Too much retrieved context can introduce noise and false confidence.
• The speaker ties failure to data modeling choices before deployment.
• The system improves by encoding what it could not previously interpret.
• The central engineering constraint is not only inference, but recall and adaptation.
• Knowledge graphs are presented as a mechanism for turning experience into structure.
• The episode frames graph-based systems as a way to manage boundaries in enterprise AI.
• The enterprise failure mode is upstream of the model itself.
• The episode favors adaptive systems that learn from common edge cases.
• Action should be grounded in knowledge to reduce harmful or misleading outputs.
• The conversation connects enterprise adoption risk to schema and terminology gaps.
• The learning loop depends on translating user questions into shared terminology.
• The system must remember enough to compare current behavior with prior behavior.
• The episode implicitly rejects a purely prompt-based approach to enterprise agents.
• The graph acts as a governed memory of vocabulary and concept mapping.
• The discussion suggests that successful agents need both restraint and learning.
• Operationally, the key is not scale alone but the quality of the encoded relationships.
• The episode portrays knowledge graphs as a mechanism for organizational learning.
• The failure of enterprise AI is attributed to preparedness and data modeling.
• The agent should build expertise from repeated interaction rather than reset on each turn.
• A practical graph update loop can capture recurring terminology from failed queries.
• The episode's strategic message is that knowing is the prerequisite for doing in AI.
• The technical thesis centers on memory, context control, and graph-backed adaptation.
• For investors and operators, the implication is that winning AI systems will encode organizational knowledge, not just call models.
• The conversation emphasizes feedback-driven improvement over static deployment.
• A useful agent needs durable state, curated context, and a mechanism to learn missing terms.