Build Daily
Tinley Park · October 13, 2026Managed Deep Agents | Deploying mission critical agents to production
Victor Moreira, Product Manager at LangChain, introduces Managed Deep Agents (now in public beta), which bundles the Deep Agents harness with managed LangSmith infrastructure so teams can define an agent as a folder, test locally, and deploy via the mda CLI. Using LangChain's internal go-to-market agent as a running example, he walks through production concerns—caller identity, scoped credentials, channel-aware memory, versioned context, and evals—and shares case studies from Madrigal and Consensus before closing with the 0.8 release highlights.
The outline
Tap a timestamp to jump there in the video.
What is an agent harness, and why does context matter?
Victor introduces the concept of an agent harness and explains why context is central to building effective agents.
The four core areas that make up a deep agent
He breaks a deep agent down into four core areas that together define its behavior and capabilities.
Madrigal spins up research agents in hours instead of weeks
Madrigal's team built a pharmaceutical research agent with Deep Agents on LangSmith, cutting build time from weeks to hours.
What changes when a local agent goes to production?
The talk examines what changes when a local agent prototype must go to production, surfacing identity, credential, and observability gaps.
How Managed Deep Agents bundles harness and infrastructure
Managed Deep Agents is presented as a unified product that bundles the Deep Agents harness with managed LangSmith infrastructure for end-to-end deployment.
An agent as a folder, shipped with three mda commands
An agent is defined as a folder of instructions, skills, and tools, and shipped to production with three mda CLI commands.
How LangChain's GTM agent maps onto files
LangChain's internal go-to-market agent is used as a concrete example, mapping its components onto the folder-based file structure.
Who is asking, and whose account should the agent use?
The session addresses caller identity across Slack and web surfaces and how user- and agent-scoped credentials can pause a task mid-run for user authentication.
Which memories can the agent use in a DM or shared channel?
Memory access control is discussed, distinguishing what an agent can recall in a direct message versus a shared channel.
Versioned context in Context Hub and evals with Harbor
Versioned context is managed through Context Hub, and evaluations are run with Harbor and traced back to LangSmith experiments.
How Consensus builds agents for different parts of its org
Consensus is shown building separate agents tailored to different parts of its organization using the Managed Deep Agents framework.
What's new in the Managed Deep Agents 0.8 release
The 0.8 release adds HTTP channels, user-scoped memory, Slack file transfers, and built-in search powered by Parallel.
The production checklist LangChain walks through—scoped credentials, versioned context, channel-aware memory, evals—is the part a solo builder skips to ship fast, then pays for in late-night debugging. Bundling that into a CLI deploy from a local folder is a reasonable convenience, but the harder question for a one-person shop is whether caller identity and channel-aware memory matter before the second user arrives.