Overview of If context is king, architecture is the castle
In this Stack Overflow Podcast episode recorded live at the AI Agent Conference in New York, Ryan Donovan speaks with Matt DeBergalis, CEO of Apollo GraphQL, about why context is becoming the central problem in AI agents—and why architecture, not just model capability, will determine whether enterprise AI succeeds. The conversation argues that agents need more than raw access to APIs: they need semantics, structure, governance, discovery, and strong security boundaries. DeBergalis makes the case that GraphQL, MCP, and principled platform architecture are key building blocks for the next wave of agentic software.
Key themes and takeaways
Context is more than data access
- AI agents need meaning, not just connectivity.
- Useful context includes:
- what data represents
- how systems relate to one another
- business rules and intent
- metadata, typing, and governance
- The core argument: agents can’t operate well if they only see “the wires” and not the semantics behind them.
GraphQL as a model for agent-friendly architecture
- Apollo frames GraphQL as a proven way to expose structured, introspectable APIs.
- GraphQL already solves a related problem for humans and apps: asking for exactly the fields needed.
- That same pattern maps well to AI:
- reduces excess context
- improves reasoning quality
- supports discovery
- enables better security and cost control
The enterprise AI shift is organizational, not just technical
- Adopting AI-first workflows is described as a major organizational change.
- AI will affect not just engineers, but also non-technical employees.
- Companies will need to:
- redesign workflows
- codify informal processes
- create new operating norms for human-agent collaboration
Good architecture matters more as systems get more autonomous
- DeBergalis emphasizes that the right answer is not “use AI everywhere,” but draw clear boundaries:
- deterministic vs. non-deterministic systems
- where models should be trusted
- where human-controlled infrastructure and security must remain in place
- He repeatedly returns to the idea that standards, contracts, and stable architecture are what make fast-moving AI systems manageable.
Security, governance, and cost concerns
Security is the biggest near-term risk
- The transcript stresses that agents are effectively untrusted, adversarial software unless carefully constrained.
- Enterprises cannot allow agents broad access without strong controls because of:
- data exfiltration risk
- insider-threat scenarios
- East-West security challenges inside the organization
- Traditional North-South API security is not enough; agentic systems require better internal governance.
Context management also solves for token cost
- The episode draws a parallel between GraphQL field selection and prompt/context management:
- only send what is needed
- avoid unnecessary token burn
- keep reasoning quality high
- Context is framed as both a security problem and a cost problem.
- Still, DeBergalis argues that organizations should not over-optimize too early at the expense of finding what works.
Measurement and closed-loop systems matter
- Successful AI systems need:
- observability
- metering
- feedback loops
- proof that the system is working
- “Closed-loop” operation is presented as essential for enterprise trust.
MCP, GraphQL, and the future of agent tooling
Why MCP matters
- MCP is described as an important standard for letting agents interact with tools and services in a more flexible way.
- It complements GraphQL well:
- MCP handles agent interaction patterns
- GraphQL provides structured querying and discovery
Discovery at scale will be the hard problem
- The world will not have one perfect catalog of everything.
- Companies will need dynamic discovery across:
- internal systems
- SaaS tools
- third-party/public resources
- DeBergalis suggests search, introspection, and routing will be central to making this work.
The future is many more software surfaces
- He predicts a rise in:
- ad hoc software
- on-demand dashboards
- agent-created workflows
- microservices written or augmented by agents
- This creates a new platform challenge: how to find, govern, and secure the right capability at the right time.
Practical advice for platform teams
Think architecturally, not just tactically
- Don’t be swayed by marketing claims from AI vendors.
- Evaluate systems based on:
- open standards
- interoperability
- security boundaries
- long-term maintainability
- clear responsibility lines
Preserve agility while protecting foundations
- Platform teams should separate:
- what must remain stable
- what can evolve with the models
- The episode warns against overcommitting to assumptions about current model capabilities, since they can change quickly.
Expect more budget, but spend wisely
- AI adoption will likely increase infrastructure and platform spend.
- The goal is not zero cost; it’s controlled experimentation with strong guardrails.
- Early exploration may require some inefficiency to discover the right patterns.
Notable insights
- “Agents need semantics.” They need more than API access; they need intent, metadata, and structure.
- “Context is king, architecture is the castle.” The episode’s core thesis is that context only works when built on solid systems design.
- “You’re going to need a strategy for adaptation.” AI will change, but enterprises can’t throw away existing systems—they need a durable way to evolve.
- “Look past the marketing copy… and look to the architecture.” The strongest advice in the conversation.
Resources mentioned
- Apollo GraphQL: apollo.dev
- MCP: referenced as an emerging standard and discussed in relation to the Linux Foundation
- GraphQL: positioned as a foundational standard for context-aware, agent-ready systems
Bottom line
This episode argues that enterprise AI success will depend less on hype around models and more on how organizations structure access, context, discovery, and security. Apollo’s view is that GraphQL-style architecture, combined with standards like MCP, can provide the principled middle layer agents need to safely and effectively operate in complex enterprise environments.