Uploaded December 2025 | Updated September 2026, 1 day ago
AI agents don’t fail in production because they’re not capable enough.
They fail because they can’t control context at scale.
In this talk, Reilly Wood, Staff Engineer at Datadog, shares a real production lesson from building Datadog’s MCP (Model Context Protocol) server — and why early agent tooling broke down under enterprise data volumes.
Agents tasked with incident investigation were:
• Pulling too much raw data
• Blowing past context limits
• Guessing from partial samples
• Producing unreliable and costly results
The solution wasn’t a larger model.
It was better structure.
By introducing SQL-based tools behind MCP, Datadog gave agents a constrained, expressive way to:
• Aggregate instead of sample
• Select only the columns and rows that matter
• Precisely control what enters the context window
• Dramatically reduce token usage and evaluation cost
• Improve accuracy and reliability in real investigations
What this reveals about enterprise AI
Reilly’s talk highlights a broader shift happening across the AI infrastructure ecosystem:
• Unstructured tools don’t scale in enterprise environments
• Giving agents shell access isn’t governance — it’s risk
• MCP acts as a control plane, not just a convenience layer
• Structured interfaces (like SQL) are safety mechanisms, not limitations
This isn’t a theoretical argument. It’s a production case study from Datadog, operating at a scale where naïve agent designs simply stop working.
Where WorkOS fits in this architecture
As more teams adopt MCP and agent-driven workflows, identity becomes a foundational concern:
• Who is the agent acting on behalf of?
• What data is it authorized to access?
• How are permissions scoped across organizations and environments?
• How are actions audited and governed?
WorkOS provides the enterprise identity and access layer that enables MCP-based systems to operate securely — handling authentication, authorization, and org-aware access so agents can safely interact with production systems.
This is what enterprise AI maturity looks like:
standards over shortcuts, structure over sprawl, and infrastructure designed for scale from day one.
AI agents don’t fail in production because they’re not capable enough.
They fail because they can’t control context at scale.
In this talk, Reilly Wood, Staff Engineer at Datadog, shares a real production lesson from building Datadog’s MCP (Model Context Protocol) server — and why early agent tooling broke down under enterprise data volumes.
Agents tasked with incident investigation were:
• Pulling too much raw data
• Blowing past context limits
• Guessing from partial samples
• Producing unreliable and costly results
The solution wasn’t a larger model.
It was better structure.
By introducing SQL-based tools behind MCP, Datadog gave agents a constrained, expressive way to:
• Aggregate instead of sample
• Select only the columns and rows that matter
• Precisely control what enters the context window
• Dramatically reduce token usage and evaluation cost
• Improve accuracy and reliability in real investigations
What this reveals about enterprise AI
Reilly’s talk highlights a broader shift happening across the AI infrastructure ecosystem:
• Unstructured tools don’t scale in enterprise environments
• Giving agents shell access isn’t governance — it’s risk
• MCP acts as a control plane, not just a convenience layer
• Structured interfaces (like SQL) are safety mechanisms, not limitations
This isn’t a theoretical argument. It’s a production case study from Datadog, operating at a scale where naïve agent designs simply stop working.
Where WorkOS fits in this architecture
As more teams adopt MCP and agent-driven workflows, identity becomes a foundational concern:
• Who is the agent acting on behalf of?
• What data is it authorized to access?
• How are permissions scoped across organizations and environments?
• How are actions audited and governed?
WorkOS provides the enterprise identity and access layer that enables MCP-based systems to operate securely — handling authentication, authorization, and org-aware access so agents can safely interact with production systems.
This is what enterprise AI maturity looks like:
standards over shortcuts, structure over sprawl, and infrastructure designed for scale from day one.
