The 6 AI Architecture Primitives
Most teams begin agentic AI design with a model, prompt, and orchestration framework. That can create impressive demos—but not necessarily systems that are safe, reliable, cost-controlled, or ready to act in the real world.
This session introduces six architecture primitives that every production AI agent needs:
Reasoning — turning an objective into bounded plans and decisions
Memory — retaining the right context without carrying unsafe or stale state
Tools — taking action through APIs, workflows, and enterprise systems
Evidence — grounding outputs in current, authorized, traceable facts
Policy — defining authority, permissions, constraints, and approval rules
Control — observing behavior, enforcing budgets, degrading safely, and stopping unsafe autonomy
Participants will see how these primitives work together as an AI agent moves from “answering a question” to making a real-world recommendation or action.
An agent that has memory but no policy can retain information it should not use. An agent with tools but no authority control can take actions beyond its business mandate. An agent with strong reasoning but weak evidence can produce a convincing wrong answer. An agent with all of these capabilities but no control layer can retry itself into cost spikes, duplicate side effects, or unsafe decisions.
The key challenge is not simply building more capable agents. It is governing autonomy: deciding what the AI may know, use, recommend, commit, defer, and stop doing when conditions become uncertain.
Session details:
Using an AI-commerce agent as a live architecture case, the session traces a request from customer question to possible business action. The agent must retrieve customer and policy context, evaluate evidence, call inventory and payment tools, and decide whether to answer, recommend, issue a refund, or escalate.
Then the system is pressure-tested through Authority and Degradation simulations:
- The agent retrieves a stale return policy but a current exception exists elsewhere.
- Inventory is delayed, but the agent wants to promise two-day delivery.
A payment tool times out after a possible charge. - The model has enough information to recommend an action—but lacks authority to commit it.
- Tool calls and retries exceed the agreed churn budget.
- A required source becomes unavailable or an entitlement check fails.
- The system is operationally healthy, but its evidence is no longer trustworthy.
For each scenario, participants determine:
- Which architecture primitive has failed or become insufficient?
- What is the agent still allowed to say or do?
- Which policy and evidence checks are required?
- Should the system remain normal, become conservative, defer, freeze a capability, or escalate to human review?
- What receipt must be recorded so the decision can later be replayed and defended?
Value for participants:
- A clear mental model for designing production-grade agentic AI beyond prompts and model selection.
- A way to separate AI reasoning from business authority and policy enforcement.
- Practical patterns for memory boundaries, tool permissions, evidence qualification, and safe retrieval.
- A method to prevent uncontrolled retries, tool loops, excessive cost, and duplicate actions.
- A degradation strategy that reduces autonomy safely without shutting down the full customer experience.
- Reusable questions to pressure-test an existing AI agent design.
Attendees will leave with:
- The Six AI Architecture Primitives reference model
- An Authority Boundary Map for agent capabilities
- An Evidence and Policy Admission Checklist
- A Tool and Churn Budget template
- An Authority and Degradation Scenario exercise
- A Stop Authority Ladder: Normal → Conservative → Defer → Freeze → Human Review
About Rohit Bhardwaj
Rohit Bhardwaj is a Director of AI & Data Architecture at Salesforce, where he focuses on enterprise AI, agentic systems, cloud-native architecture, distributed systems, data platforms, security, and large-scale transformation.
Over his career, Rohit has designed and led complex enterprise platforms across AWS, Google Cloud, microservices, real-time data, API ecosystems, resilient distributed systems, and AI-enabled architectures. His work increasingly focuses on the challenges enterprises face as software evolves from deterministic services to AI-native and agentic systems—particularly around reliability, governance, evidence, security, observability, cost, and safe autonomy.
Rohit is the author of System Design with AI Interview Guide: Designing Scalable, Agentic, and Defensible Systems, published by Apress. The book presents a modern approach to system design covering scalability, distributed systems, AI architecture primitives, security, reliability, economics, agentic systems, and real-world architectures including e-commerce, ride sharing, payments, fraud detection, messaging, video streaming, file storage, and search. (Springer Link)
Book:
Amazon: https://a.co/d/09Zs1twa
Publisher / Springer Nature: https://link.springer.com/book/10.1007/979-8-8688-2782-2
O'Reilly: https://learning.oreilly.com/library/view/system-design-with/9798868827822/
Rohit is also an O’Reilly instructor and a frequent speaker at technology conferences including No Fluff Just Stuff, UberConf, GIDS, and other international events. His talks focus on practical architecture lessons from building and operating complex systems, including AI control planes, trusted agents, inference at scale, evidence-first RAG, AI security, distributed-system failure, and AI-era software architecture.
As a trusted advisor and architecture leader, Rohit works at the intersection of business strategy and deep technical architecture—helping teams translate complex business problems into scalable, resilient, secure, and economically sustainable systems.
Rohit holds an MBA in Corporate Entrepreneurship from Babson College and graduate-level education in Computer Science from Boston University and Harvard University.
Connect with Rohit:
LinkedIn: http://linkedin.com/in/rohit-bhardwaj-cloud
X / Twitter: @rbhardwaj1