Building an AI agent is getting easier.
Running one reliably in production is still the hard part.
At Interrupt 2026, LangChain introduced a major set of updates focused on production agent infrastructure. Since then, the platform has expanded further with LangSmith Engine v2, public beta access for Managed Deep Agents, and additional governance capabilities.
Here’s the current stack developers should know about.
The original LangSmith Engine was introduced at Interrupt 2026 to help teams analyze production traces, identify recurring failures, diagnose root causes, and propose fixes.
Engine v2, released in September 2026, goes further:
🔴 Proactively red-teams agents to identify potential issues before they appear in production.
📉 Detects inefficient agent trajectories and trends in error rate, latency, and cost.
🧪 Automatically tests proposed prompt and code fixes before human review.
🔄 Reproduces failures, tests fixes, evaluates the results, and then surfaces validated changes for review.
That makes the development loop much more automated:
> Production traces
> ↓
> Issue detection
> ↓
> Root-cause analysis
> ↓
> Proposed fix
> ↓
> Automated validation
> ↓
> Human review
Agent traces are becoming much larger and more complex than traditional application logs.
SmithDB is LangChain’s purpose-built database infrastructure for agent observability. It is designed for workloads involving deeply nested spans, long-running operations, full-text search, JSON filtering, and trace reconstruction.
LangChain reports up to 15× faster performance on core LangSmith workloads compared with its previous infrastructure.
Running long-lived autonomous agents yourself means managing:
Runtime infrastructure
Persistence
Checkpointing
Memory
Tool execution
Sandboxes
Streaming
Observability
Managed Deep Agents moves much of that operational layer into a hosted LangSmith runtime.
Developers can define agents using the Deep Agents framework and deploy them without building and maintaining their own agent server infrastructure. Managed Deep Agents became available in public beta in August 2026.
Giving an AI agent the ability to execute code introduces obvious security concerns.
LangSmith Sandboxes, which reached GA at Interrupt 2026, provide isolated execution environments for workloads involving generated code, shell commands, files, dependencies, and data analysis.
The environments use hardware-virtualized microVMs for stronger isolation from your main infrastructure.
Production agents also need controls around cost, data, permissions, and context.
LLM Gateway
LangSmith’s LLM Gateway sits between agents and model providers and provides controls such as:
Spend limits
Rate limits
Model fallbacks
PII and secret redaction
Centralized usage visibility
The Gateway entered public beta in July 2026.
Context Hub
Context Hub provides a centralized way to manage and version the files and information that shape agent behavior, including instructions, policies, examples, and skills.
Messages View
For complex agent workflows, Messages View makes multi-turn traces easier to read and understand, reducing the friction of debugging long agent sessions.
🔍 Why This Matters
The interesting part isn't any single feature.
The bigger story is that agent infrastructure is becoming a full production stack:
AI Models
↓
Agent Runtime
↓
Tools + Memory + Files
↓
Secure Sandbox
↓
Observability + Evaluation
↓
Governance + Cost Controls
↓
Production Scale
We're moving from “build an agent” to “operate an autonomous system.”
That changes the engineering questions too.
It's no longer only:
Which model should I use?
It's also:
How do I monitor the agent?
How do I test it?
How do I control its cost?
How do I secure tool execution?
How do I recover from failures?
How do I improve it continuously?
That is where the next generation of agent infrastructure is being built.
What’s your take? Are you moving toward managed agent runtimes, or do you prefer to keep the runtime infrastructure self-hosted?