LangChain AI Updates: November 3, 2026

1. LangSmith Deployments Get Run Priority Levels and Human-in-the-Loop Prompts Over MCP

LangChain published its LangSmith Cloud changelog for October 26 to November 2 on November 2. A priority section in langgraph.json can now name a callback that sets each thread’s priority from -5 to 5 at its first run: when every worker is busy, lower levels start first and runs at the same level start oldest first, threads return run_priority and fairness_group, and metrics report queue depth and wait time per level (ordering applies only to deployments using Redis and needs langgraph-sdk 0.4.9 and langgraph-cli 0.4.38 or later, the client-side pieces covered on October 24). Separately, when a served graph pauses, MCP clients on protocol revision 2026-07-28 that support form elicitation are now asked for input and the tool call resumes with the answers, covering text questions, interrupts with a response_schema, form and URL questions relayed from MCP tools, and LangChain human-in-the-loop reviews to approve, edit, reject or answer; clients that do not declare the elicitation capability get error -32021 naming what to enable. Together these let one deployment favor interactive threads over background work, and let an agent exposed as an MCP tool stop for a person’s input without a custom UI. Source

2. LangSmith Splits Workspace Administration Into Granular Permissions and Enforces New Ingest Limits

LangChain. The same changelog replaces the broad workspaces:manage permission with action-level permissions for settings, retention, usage controls, security policy, resource tags, integrations, sandbox proxy profiles, credentials, restricted role assignment and workspace deletion (existing custom roles holding workspaces:manage receive them automatically), and adds organization-scoped custom roles that can be assigned to members, pending invitations and service account keys. Multipart and OTel batch ingestion, and /runs/batch, now enforce all short-term usage limits with HTTP 429, and tenants still on default limits move to 300,000 events per minute, 1,000,000 tracing requests per hour and 15 GB of tracing payloads per hour. Tool Gateway can write each tool call into the caller’s trace as a child run recording the tool, server, identity, outcome and timing but never arguments or results, tracing for OpenAI’s hosted Agents API is in beta through wrap_openai, and CLI profiles can store an API key as a reference with --api-key-env or --api-key-cmd instead of in plaintext. POST /api/v1/runs/group is deprecated in favor of POST /api/v2/threads/query, and it will be removed on January 31, 2027 along with Python SDK methods such as list_runs() and read_run(), which now emit deprecation warnings. Source

3. Deep Agents Code 0.2.3 Defers MCP Tool Schemas to Provider-Native Tool Search

LangChain released deepagents-code 0.2.3 on November 2, followed later that day by 0.2.4. MCP tool schemas are now loaded on demand through the model provider’s native tool search instead of all at once, redundant MCP startup probes were removed, the official marketplace is included by default, and the CLI now requires an explicit or saved model selection. Hook command rewrites are applied before a tool approval is requested, so the user approves the command that will actually run, and auto mode is blocked when its classifier model cannot be resolved, with inherited classifier model switches disclosed. Fixes enable Claude prompt caching on OpenRouter, skip non-UTF-8 subagent AGENTS.md files instead of failing startup and correct coding-agent Git trace attribution, while 0.2.4 adds clearer messages for workspace conflicts, MCP login failures and disabled Slack MCP access. Source

4. Harrison Chase Lists Open Context Engineering Problems for Always-On Agents

LangChain CEO Harrison Chase argued in a November 2 post that builders have converged on one agent architecture (an LLM in a loop calling tools inside a sandbox) and on standard context techniques: summarizing long threads, offloading large tool results, skills for progressive disclosure, tool search, subagents for context isolation and bash as a primary tool. Proactive, always-on, goal-oriented agents, he writes, bring problems without general solutions: calibrating how often a background agent pushes notifications, deciding what a memory layer should remember, notifying a main agent when asynchronous subagents or tools finish (including while it is mid-run or paused for approval), keeping agents trying new experiments toward long-horizon goals instead of repeating themselves, designing goals agents cannot game, and managing browser context, where both the DOM and screenshots are large objects. Several of these map onto features LangChain added to Managed Deep Agents in late October, including the Schedules SDK, asynchronous subagents and browser use. Source

5. LangChain Sets a Python 3.10 Cutoff; LangChain.js Fixes Claude Reasoning Replays on Bedrock

LangChain shipped langchain-core 1.7.4 and langchain 1.4.8 on November 2 with a notice that Python 3.10 support ends in langchain-core 1.8.0 and langchain 1.5.0, making the current minor versions the last to support it; core 1.7.4 also stops the Anthropic translator from mutating message content, and langchain-anthropic 1.7.8 and 1.7.9 strip the index from compaction blocks, handle null cache-creation token counts and report streamed content-block tokens to callbacks. In LangChain.js, @langchain/aws 1.5.2 replays signed Bedrock reasoning blocks whose text is empty or missing, so models that omit the thinking summary, such as Claude Sonnet 5 with its default display: "omitted", no longer fail with HTTP 400 on the turn after a tool call, and internally created Bedrock clients now default to maxAttempts: 1 so SDK retries do not stack with LangChain retries. @langchain/mcp-adapters 2.0.2 connects to configured MCP servers concurrently during discovery, so listing tools takes as long as the slowest server rather than the sum of all, and @langchain/core 1.3.3 adds a tracePolicy to RunnableSequence that controls what is sent to LangSmith for the sequence’s own run. Source Source