Managed Deep Agents is now in public beta

Jessica Okay, this is going to sound insane on a podcast — Managed Deep Agents just hit public beta today.

Cathy Wait, really? LangSmith’s thing? I saw the tweet but figured it was another preview.

Jessica Nope, full public beta. You can author a Deep Agent in Python or TypeScript, test locally, and deploy with one command — mda deploy — and LangSmith handles the runtime.

Cathy Right. So it’s not just the harness, it’s the managed runtime wrapped around it.

Jessica Exactly. You get the open source Deep Agents harness — which is model agnostic, lets you bring your own model, tools, instructions — but LangSmith takes care of persistence, memory mounts, skill loading, sandbox lifecycle, deployment.

Cathy Yeah, that’s the part that usually eats quarters to build. Durable execution so long-running agents can pause, retry, resume without losing work; streaming so users see progress; sandboxes for isolated file work; evals via Harbor; channels for Slack/GitHub; memory that persists across threads.

Jessica And identity and auth for multi-user agents — scopes threads per end user ID from your OIDC provider so each user’s threads are isolated.

Cathy Mm-hm. So the pitch is: you spend time on the agent’s behavior — prompts, tools, middleware, identity rules, evals, domain logic — instead of rebuilding the same infra every agent needs.

Jessica Yeah, no, you’re completely right. That’s the product win — not smarter agents but visibility into where context was discarded, where a handoff happened, whether anyone checked it.

Cathy Exactly. Harness is the product. That’s been our thesis since Muse Code — the boring infrastructure is exactly what’s missing.

Jessica Remember when you swore off GraphQL? This feels similar — teams are already using it to ship faster by focusing on agent behavior instead of infra.

Cathy Oh interesting. So who’s the target user here? Is this for teams that already have agents in production or those stuck in prototype hell?

Jessica Both, really. If you’re prototyping, you can test locally with mda dev in LangSmith Studio. When you’re ready, mda deploy compiles the project, syncs deploy-owned context to Context Hub, uploads the build, creates a hosted LangSmith deployment.

Cathy Got it. And the harness itself is open source — you can own it, fork it, extend it. Managed Deep Agents just gives you the managed runtime option.

Jessica Exactly. Deep Agents is that reusable harness companies can own and control. Managed Deep Agents helps you take it to production without managing the underlying infrastructure yourself.

Cathy Right. So it’s not vendor lock-in on the agent logic — you keep that. It’s lock-in on the runtime, but you’re trading undifferentiated infra work for a managed service.

Jessica Yeah, and given how much teams struggle with sandbox provisioning, lifecycle, cleanup, durable threads, persistence, cancellation, retry behavior — that trade’s worth it for a lot of folks.

Cathy Sure. Though I’ll be watching to see if the managed runtime introduces new limits — cold starts, scaling quirks, vendor-specific gotchas that only show up at scale.

Jessica Valid point. Public beta’s the time to find those edges. Though LangSmith’s built on the same Deployment Agent Server teams already use to run agents in production, so hopefully it’s battle-tested.

Cathy Yeah, I see. So it’s an opinionated runtime for Deep Agents built on LangSmith Deployment — packages the operational patterns required for product agents.

Jessica Exactly. Sandboxes for code execution and filesystem work, Harbor evals without the setup work, channels that let your agent receive events from Slack and respond without a separate integration service.

Cathy That’s genuinely useful. Especially for agents that collaborate with users — like a code review agent that can comment on GitHub or a support agent that responds in Slack.

Jessica Yeah, users can tag an agent where the team is already discussing the work. No more context switching.

Cathy Okay, that’s good. So real user story: team builds a coding agent, defines a Slack channel file, gets real-time feedback in their existing workflow instead of jumping between tools.

Jessica Precisely. And because memory persists across threads via Context Hub, the agent can carry durable preferences and context across conversations — redeploy without wiping what the agent learned.

Cathy Right. So the evals piece — mda evals init, mda evals compile — packages your agent for Harbor so you can test behavior, tool use, state changes before and after deployment.

Jessica Yes. Every run is traced in LangSmith, so production failures become future test cases, closing the feedback loop.

Cathy That’s the kind of boring infrastructure that actually makes agent work adoptable. Annoyingly, yes.

Jessica Ha! okay. I knew you’d come around on the infra thesis eventually.

Cathy Stop it—. I’m just saying the boring controls win — they’re load-bearing.

Jessica That’s such an Exploring Next sentence. Anyway, Build Next: if you want to try it, the commands are in the blog — uv tool install managed-deepagents for Python or npm install -g managed-deepagents for TypeScript, then mda init research-assistant, and so on.

Cathy Got it. I’ll spin up a test agent this weekend and see how the sandbox story feels in practice.

Jessica Nice. Keep me posted — and if you hit any rough edges, that’s exactly what the beta’s for.

Cathy Will do. Though I reserve the right to complain about cold starts if they suck.

Jessica Deal. Signing off — thanks for the chat, Cathy.

Cathy Anytime, Jessica. This was actually fun.