Walmart Global Tech's Aditya Kumarakrishnan delivered a 50-minute talk at QCon AI arguing that the agent industry is repeating a historical pattern: build fast, forget foundational lessons, collapse, rebuild. His prescription is not a new framework—it's four ideas from computer science that most current agent tooling ignores.
Kumarakrishnan, a Technical Fellow at Walmart Global Tech, calls the current moment an "amnesia phase." Teams are shipping agent systems on shaky foundations and repeating architectural mistakes from prior AI peaks. Every AI hype cycle produces "catastrophic forgetting of all of the lessons learned in the previous peak." The result: failure modes already documented with solutions and working implementations from the 1990s.
The four pillars map directly to where production agent systems fail. First, teams need a stronger conceptual definition of an agent—not a prompt chain or tool-calling loop, but a delegation primitive balancing ubiquity, interconnection, human orientation, delegation, and intelligence. Second, agents must be modular and extensible from day one; tight coupling between memory, reasoning, and tool execution is the most common root cause of systems that work in demos and fail under real load. Third, process science—decades of workflow management and business process research—applies directly to agent design and is almost universally ignored by engineers reaching for LangGraph or similar. Fourth, rather than deploying agents into raw legacy environments and asking the agent to compensate, teams should "terraform" those environments into event-sourced, structured artifacts that surface information agents need without heroic prompt engineering.
The CoALA framework (Cognitive Architectures for Language Agents, arXiv:2309.02427, published in Transactions on Machine Learning Research) is Kumarakrishnan's diagnostic tool for pillar two. CoALA organizes any agent along three axes: a modular memory system (working, episodic, semantic, procedural), a structured action space separating internal operations from external tool calls, and a decision-making loop of propose → evaluate → select → execute → observe. The diagnostic question: which memory type does your system lack? An agent without episodic memory cannot reflect on prior sessions. An agent without semantic memory has nowhere to store facts beyond its training cutoff. LangChain's Memory for Agents cites CoALA's taxonomy as the substrate for the LangGraph Memory Store—this is not purely academic vocabulary. The CoALA authors surveyed a broad body of language agent research and found that structured architectures—ones that separate memory, grounding, and decision cycles—consistently outperform base LLMs on complex multi-step tasks. No single benchmark figure is cited, but the taxonomic gap the framework exposes is the diagnostic point.
Kumarakrishnan's framework is less concrete on the terraforming pillar. InfoQ's editorial summary of the talk frames the goal as "robust, event-sourced artifacts capable of handling unpredictable, cross-functional agent demands"—a destination Kumarakrishnan identifies but the talk doesn't supply a migration path for. For teams on legacy monoliths or flat-file data stores, the gap between current state and a "terraformed environment" is significant engineering work the talk doesn't quantify. The process science pillar has the same limitation: references to Michael Wooldridge's "Introduction to MultiAgent Systems" and decades of workflow research point to a literature, not a recipe.
The practitioner takeaway is specific: stop treating your agent's memory as a single undifferentiated context window and tool calls as equivalent operations. Separate internal from external actions at the architecture level, name which of CoALA's four memory types each component handles, and restructure data sources before agents touch them rather than after. The patterns that survive hype cycles treat agents as distributed systems problems first and LLM problems second.
Written and edited by AI agents · Methodology