Skip to main content

Conference Brief

Executive Summary​

LangChain Interrupt 2026 (San Francisco, May 13–14, 2026) marked the field's shift from building agents to operating them. The unifying concept was the deep agent — an agent harness that wraps a model with an execution environment (virtual file system through full code sandbox), context management (skills, memory, summarization, prompt caching), human-in-the-loop steering, and delegation to subagents.

LangChain's launches reflected the same shift from control toward closed-loop operations: Managed Deep Agents, LangSmith Engine (an ambient agent that diagnoses production traces and proposes fixes as one-click GitHub PRs), LangSmith Sandboxes (GA), Context Hub, an LLM Gateway, and SmithDB — a Rust-based observability database that made trace workloads 6–15× faster. The platform now serves 100M+ agent runs and 150M+ traces per week.

Across two days, production case studies from Cisco, Toyota, Box, MongoDB, Abridge, Bridgewater, Coinbase, LinkedIn, Lyft, and Chime showed agents shipping in finance, healthcare, transportation, and the enterprise — with evals, MCP, and skills as the connective tissue. Andrew Ng and the closing "Return of the Data Scientist" session grounded it all in human judgment: redesign workflows, chase growth, and always look at your data.

Key Themes​

  1. Deep agents as the organizing abstraction. The agent harness — execution environment, context management, steering, subagents — replaced "just a model in a loop."
  2. Skills as a shared primitive. Progressively disclosed, reusable, portable knowledge bundles — not framework lock-in.
  3. Sandboxes for untrusted code. Agents write and run real code; sandboxes (sub-second spin-up, egress proxy, snapshot/fork) make that safe.
  4. From observability to action. LangSmith Engine turns traces into clustered issues and proposed fixes — an agent that improves agents.
  5. Evals are infrastructure. Written by engineers and legal/compliance, run online and offline, and increasingly generated by agents themselves.
  6. MCP as the integration layer. Swap backends without touching agents; the standard matured from "emerging" to "assumed."
  7. Open memory and context standards. agent.md, skills, and memory should remain portable.
  8. Production at scale. 100M+ runs, 50,000+ agents at Toyota, 14M at ElevenLabs, 250 health systems at Abridge.
  9. Coding agents lead; knowledge work follows. Verifiable output and open permissions make coding the beachhead; everything else needs change management.
  10. Stay model- and framework-agnostic. Cost and rapid model churn reward portability and optionality.
  11. The data scientist returns. Experiment, measure, improve — and always look at your data.

See the full narrative in the Conference Recap and every session in the Presentation Directory.