LangChain CEO Shares Framework Updates for AI Agents

Dynamic subagents in deepagents!

Lets you spin up subagents programmatically

We highlight 6 diff use cases for this

this blog covers 6 common workflow patterns we see with dynamic subagents (fanout and synthesize, adversarial verification, etc)

if you'd prefer to learn w/ a video, check out this excellent guide from @colifran_ with examples and trace walk throughs!

Want to learn more? Here's a great demo from @colifran_ https://www.youtube.com/watch?v=5AkdMangfNk

Deep Agents now supports dynamic subagents. Instead of invoking subagents with tool calls, the main agent writes orchestration code to coordinate work at scale.

This enables workflows like processing hundreds of documents or analyzing thousands of datapoints with deterministic coverage.

6 common orchestration patterns and how to get started today ⤵️

@sydneyrunkle this is so clean 😭

we just released dynamic subagents, which let your agent programmatically orchestrate subagents in a code interpreter.

this lets agents do work at scale that tool calls can't reliably handle, like processing hundreds of documents or analyzing thousands of datapoints!

Masterclass deep-dive into programmatic spawning of subagents - highly encourage taking a look! @colifran_ dropping elite ball knowledge per usual

@sydneyrunkle Any eval dynamic agents project

@sydneyrunkle Doesn't this explode the usage cost in most cases? I have seen so many people not blow their own foot off with deep agents because they asked 1 agent per page, try to launch 500k pages and the orchestrator only spawns like 4

@LangChain The orchestration is the easy 20%. Running a fleet of agents myself, the hard part is the verification layer: which subagent output do you trust enough to act on without reading it.

Coordinating agents at scale is great until one is confidently wrong. How are you handling that?

@LangChain makes debugging feel way more natural actually

does it still stream back what each subagent did or is it blackbox now?

@hwchase17 The interesting part is less “more agents” and more the control contract between them.

For production I’d want each subagent to return: scoped context, tool calls + cost, failure mode, handoff artifact, verifier result.

Otherwise parallelism just multiplies invisible state.

@hwchase17 Interesting — programmatic subagent orchestration could unblock a lot of parallel task workflows. Curious which use case sees the highest compute efficiency gain vs. manual routing.

@LangChain Good for documents. How about orchestrating production workflows?

@LangChain @colifran_ This is crucial for production AI agents. Dynamic allocation = self-healing architecture.

@hwchase17 wait, are you using these for data processing tasks?

@sydneyrunkle Really sharp approach. Letting the agent write orchestration code instead of chaining tool calls should give much more reliable scaling and deterministic coverage on big tasks. Excited to see this in practice

@LangChain @colifran_ What happens when a subagent returns corrupted data instead of crashing? Does the system have validation layers to catch that?

@LangChain For enterprise, the fan-out/fan-in pattern is likely the highest ROI—parallel processing of documents or data at scale is where most teams need help today.

@LangChain The orchestration approach unlocks true compound reasoning—chains of specialized subagents handling complex, multi-step enterprise workflows with proper context isolation.