LangSmith ships Engine v2 with red teaming and agent fine-tuning

25 Sep 2026 (ET): LangChain announced LangSmith Engine v2 at Interrupt NYC — red teaming, broader issue detection, and automated fix validation — plus Managed Deep Agents 0.8 (user-level memory and built-in web search), Trajectories (readable agent-session views), Fine-Tuning via smithtune, and Custom Apps for workspace UIs.

SOFTWARE desk — an always-on agent in Slack or on a customer channel fails in ways a person cannot watch turn by turn, and the session log is the only record of the path it took. LangSmith’s announcement puts an in-platform agent on that record so it can hunt failures before a user hits them, and it turns the same readable session into examples for fine-tuning a model on the company’s own work. The hunt and the automatic fix check are a private beta for teams already on LangSmith Deployment, and the two company posts do not use the same trace count.

What Engine is, and whose trace count that is. Engine is an in-platform agent that analyzes production traces and works through the steps of fixing an agent: spot the issue, find a cause, propose a fix, test it, ship it, and watch for a regression. A trace, here, is the recorded path of one agent run, including the messages and the tool calls. That gloss is this desk’s. The overview says that since the launch in May, Engine has helped engineers analyze more than 60 million traces and diagnose tens of thousands of issues. The Engine v2 detail post, by Ben Tannyhill and Trammell Saltzgaber, says Engine has analyzed more than 70 million traces and diagnosed tens of thousands of issues. Helped analyze, and has analyzed, are two different verbs. 60 million is the 25 Sep overview. 70 million is the 24 Sep detail post. Do not collapse them into one number. This desk did not count the traces.

What the detail post’s clock is, kept apart from the overview. The visible byline is September 24, 2026. That line does not print an hour. schema.org datePublished is 2026-09-24T14:04:00.000Z, which is 2:04 p.m. UTC and 10:04 a.m. Eastern. dateModified is 2026-09-24T17:22:02.282Z, which is 5:22:02 p.m. UTC and 1:22:02 p.m. Eastern, three hours, eighteen minutes, two seconds, and 282 milliseconds later. Those conversions are this desk’s. This desk read the page as it stood. It did not diff those three hours. The 24 Sep post is the Engine v2 detail. It is not a second product.

Red teaming, and which page prints the private beta. The overview says Engine can use production traces and the agent’s code repository to generate hypotheses about issues that may not have appeared in production yet, test those hypotheses, and surface confirmed failures for review. A hypothesis, here, is a guessed failure the tool then tries to confirm. That gloss is this desk’s. The detail post says Red Teaming looks for issues such as hallucinations and violations of system prompts before they affect users. A hallucination, here, is an answer the agent states that is not backed by the material it was given. A system prompt is the standing instruction the builder wrote for the agent. Those glosses are this desk’s. The detail post says Red Teaming is available in Private Beta to existing LangSmith Deployment users, and that a team can apply to join. The overview’s red-teaming section this desk read does not print Private Beta. Do not move that label onto the overview.

The harder-to-spot issues, and whose list that is. Both posts say Engine v2 looks past outright errors and unmet requests. It flags inefficient agent paths, such as repetitive tool calls or a trajectory that runs longer than the task needs, and it tracks trends in error rate, latency, and cost. Latency is how long the answer takes. Cost, here, is the model bill for that work. Those glosses are this desk’s. The detail post adds an incorrect tool call, and a proposed more efficient alternative, and it says every issue lands in the same queue. The overview’s phrase is unnecessarily long trajectories. The detail post’s phrase is an inefficient path. Do not treat those as two different products.

How a proposed fix gets tested, and which page says private beta. For agents on LangSmith Deployment, the overview says Engine runs the offending inputs against the agent to confirm the issue, then tests candidate fixes against a broader eval set until it finds one that resolves the issue. An eval set is a saved batch of examples used to score the agent. That gloss is this desk’s. Once the fix is satisfactory, the user can open a pull request with one click. A pull request is a proposed code change waiting for review. That gloss is this desk’s. The detail post’s key line says Engine tests and validates prompt and code fixes before showing them. Its body says Engine first reproduces the failure in Deployment, then proposes a change, tests it against the same inputs, evaluates the result, and adjusts the fix. Broader eval set is the overview’s. The same inputs, then adjust, is the detail post’s. Do not collapse those two test descriptions. The detail post says this automatic check is in Private Beta for existing LangSmith Deployment users. The overview’s fix section this desk read does not print Private Beta.

Who can turn Engine v2 on, and what is still a plan. The detail post says Engine v2 is available in SaaS for LangSmith Plus and Enterprise plans. SaaS, here, means LangChain hosts it. That gloss is this desk’s. Support for Engine v2 on self-hosted plans is coming shortly. A Plus plan can enable Engine in a few clicks. An Enterprise plan reaches the account team. The post links https://www.langchain.com/pricing. This desk did not copy a dollar price. The overview says the next release of self-hosted LangSmith will support bring-your-own-key for Engine. The detail post says that key lets a team use its own model API key for Engine’s inference while the trace data stays inside the team’s own network. Bring-your-own-key, and shortly, and next release, are three different tenses. Do not file self-hosted Engine v2 as shipped today.

The older score the detail post points at, which is not a new measurement in this package. The detail post says Engine is now more than twice as good at detecting issues, as measured by IssueBench, and that its fixes are 25 percent more effective, as measured by Terminal-Bench, “as we shared in August.” IssueBench and Terminal-Bench are the names on that sentence. Twice as good, and 25 percent, are that August comparison as the 24 Sep post repeats it. They are not a score this desk ran, and they are not the fine-tuning table below. The post links an August Engine note. This filing does not re-file that note as a second announcement.

Managed Deep Agents 0.8, and what the new memory is for. The overview and the 24 Sep post by Nathan Drezner and Karthic Subramanian describe Managed Deep Agents as the Deep Agents harness plus the infrastructure to run it. A harness, here, is the surrounding software that lets an agent keep working, not the model itself. That gloss is this desk’s. The product already stored durable agent memory, shared by everyone on a deployment. Version 0.8 adds user-level memory, scoped to the authenticated person, for preferences and other caller-specific context, kept separate from that shared agent memory. The runtime does not copy content between the two layers. Access rules can be set on each layer. The detail post’s defaults: a Slack one-to-one direct message allows both agent memory and user memory; a Slack channel or group direct message allows agent memory and denies user memory; an HTTP run allows agent memory and denies user memory. The point of that table is that a person’s private notes do not spill into a group thread. Those defaults are the 24 Sep post’s. The overview says the layers stay separate. It does not print this table. Agent memory is mounted at /memories/agent/. User memory is mounted at /memories/user/ and keyed to the caller. Those paths are the detail post’s.

Credentials, channels, and the built-in search. Managed Deep Agents now supports credentials owned by the agent and credentials owned by the user, for connections such as GitHub and Notion. Agent-owned credentials are shared. User-owned credentials matter when each person has different permissions, which the detail post names for GitHub, Linear, and Notion. The detail post says LangSmith already manages authorization for 23 services, including Linear, GitHub, and Google Workspace. 23 is that page’s count. This desk did not open all 23. Slack can now carry files: logs, spreadsheets, contracts, screenshots, and other documents. A new HTTP channel accepts a JSON webhook, so a product that can send a web request can reach the agent. JSON, here, is a plain data format. A webhook is a message one system sends to another when something happens. Those glosses are this desk’s. Web search is built in and powered by Parallel. The overview says a team does not need a separate vendor account, API key, or custom tool. The detail post says Parallel web search is free through Managed Deep Agents while that product is in beta. Free while in beta is not a promise that the meter stays off later. Quotes on the detail post, not interviews this desk conducted: Zahid, chief technology officer at Kyth.ai, on client confidentiality with user-level memory; Derek Gilbert, engineering at Consensus, on Slack and always-on triage. Those are their sentences on that page.

What the Managed Deep Agents post’s clock is. The visible byline is September 24, 2026, Nathan Drezner and Karthic Subramanian. That line does not print an hour. schema.org datePublished is 2026-09-24T14:03:00.000Z, which is 2:03 p.m. UTC and 10:03 a.m. Eastern. dateModified is 2026-09-24T15:43:10.658Z, which is 3:43:10 p.m. UTC and 11:43:10 a.m. Eastern, one hour, forty minutes, ten seconds, and 658 milliseconds later. Those conversions are this desk’s. This desk read the page as it stood. It did not diff that hour. It is the 0.8 detail, not a second launch.

Trajectories, and who they are for. A trajectory is a chronological, conversational view of one agent session. It gathers messages from the human, the AI, and the tools, across the main agent and any subagents, and shows each message once, in the order it first appeared. A subagent is a helper agent the main one calls. That gloss is this desk’s. The overview says the full trace can be too long and too nested when a person only needs the path. The 24 Sep post by Winston Huynh and Bonnie Pecevich says a subject-matter expert can score, flag, and annotate that readable view without parsing execution metadata. A subject-matter expert, here, is the person who knows the work, who may not be the person who built the agent. That gloss is this desk’s. Online evaluators can score a trajectory. The post says teams can route trajectories to annotation queues or datasets. It says the view works for traces from LangChain, LangGraph, and Deep Agents, from agent kits such as OpenAI and Claude, and from coding agents such as Codex, Claude Code, and Cursor. Those names are the post’s list of trace sources. This desk did not send a trace from each one. The get-started line says Trajectories are available now on all plans in the United States. All plans in the United States is that post’s. The overview does not print that limit.

What the Trajectories post’s clock is. The visible byline is September 24, 2026. That line does not print an hour. schema.org datePublished is 2026-09-24T14:01:00.000Z, which is 2:01 p.m. UTC and 10:01 a.m. Eastern. dateModified is 2026-09-24T15:48:07.007Z, which is 3:48:07 p.m. UTC and 11:48:07 a.m. Eastern, one hour, forty-seven minutes, seven seconds, and 7 milliseconds later. Those conversions are this desk’s. This desk read the page as it stood. It did not diff that hour.

Fine-tuning, and whose scores those are. LangSmith Fine-Tuning is a way to turn trajectories into training examples for a narrower model. Supervised fine-tuning, the method the posts name, means the model practices on examples of good behavior and updates its weights to imitate them. Weights are the numbers inside the model. That gloss is this desk’s. The command-line tool is smithtune. The overview says it builds datasets from trajectories, trains through Baseten or Fireworks, evaluates in LangSmith, and serves the tuned model. The 24 Sep post by Ankush Gola, Jake Broekhuizen, and Vivek Trivedy says Fine-Tuning is in Public Beta. A team needs a LangSmith account with traces, an API key for Fireworks or Baseten, and the smithtune tool. The post says the tool pulls trajectories, helps label which ones are good, stores that set back in LangSmith, trains, compares the tuned model with the base model, and can deploy it. Public Beta is that post’s. The overview does not print Public Beta. Quotes on the page, not interviews this desk conducted: Pranav Jain, product lead at Fireworks, on moving from curated traces to a served model; Aaron Ellis-Bloor, applied researcher at Baseten, on going from data to a training experiment in minutes.

The internal scores, kept as LangChain’s own test. On a subset of IssueBench, their internal benchmark for finding and grouping issues, a fine-tuned Kimi K3 scored 96.0. The base Kimi K3 scored 90.0. GPT-5.6 Sol scored 87.0. The post says the base Kimi model was already strong, and that harness work had stopped moving it or GPT-5.6 Sol further. On an internal set of real pull requests for a code-review agent, Qwen-3.8-27B moved from an F1 of 48.9 percent to 53.7 percent after the same kind of training. F1 is one score that balances how many flagged bugs were real and how many real bugs were found. Precision rose from 62.9 percent to 81.5 percent. Recall stayed at 40.0 percent. Model calls per review fell from 55.9 to 39.2, which the post calls 29.8 percent fewer. Tool requests fell from 65.8 to 46.5, which the post calls 29.4 percent fewer. An earlier, less selective training set reduced the F1 score. Those tables are LangChain’s internal runs. This desk did not rerun them. They are not the August “twice as good” line on the Engine v2 post.

What the Fine-Tuning post’s clock is. The visible byline is September 24, 2026. That line does not print an hour. schema.org datePublished is 2026-09-24T14:00:00.000Z, which is 2:00 p.m. UTC and 10:00 a.m. Eastern. dateModified is 2026-09-24T15:50:36.104Z, which is 3:50:36 p.m. UTC and 11:50:36 a.m. Eastern, one hour, fifty minutes, thirty-six seconds, and 104 milliseconds later. Those conversions are this desk’s. This desk read the page as it stood. It did not diff that hour. The GitHub link on the page is https://github.com/langchain-ai/smithtune. This desk did not install the tool.

Custom Apps, which this package describes on the overview and not on a separate post this desk loaded. Custom Apps let a team build and publish its own LangSmith interface inside the workspace, for jobs such as annotation queues, experiment reviews, and trace review. The overview says a team can build against LangSmith’s interfaces using templates, in-app chat, or coding-agent guidance, instead of exporting the data to a separate internal tool. Annotation, here, means a person marks whether an agent session was good. An experiment review compares outputs across versions. Those glosses are this desk’s. This desk did not publish an app.

What the card shows. The card is LangChain’s LangSmith Engine v2 launch graphic. The still is a dark panel titled LangSmith Engine v2, with three labeled blocks for red teaming, harder-to-spot issues, and automatically tested fixes. Those labels are the graphic’s. The 60 million and 70 million trace counts, the private-beta line, the fine-tuning scores, and the Interrupt date are not printed as captions on this card. The card has no desk date laid over it. The catalog chip is SOFTWARE.

Plain English for the rest of the card. LangSmith is LangChain’s product for watching and improving agents. Engine is an agent inside that product that reads production traces. Version 2 can try to find failures before a user hits them, flag wasteful paths and drifting cost or latency, and, for agents already on LangSmith Deployment, test a proposed fix before a person opens a pull request. That last step, and the red team, are a private beta on the detail post. Managed Deep Agents 0.8 gives each signed-in person a memory that stays out of the shared agent memory, lets files move through Slack, accepts a webhook, and includes web search from Parallel while the product is in beta. Trajectories are the readable session view, on all plans in the United States. smithtune turns those sessions into fine-tuning data and trains through Baseten or Fireworks. The 96.0, the 53.7 percent, and the call reductions are LangChain’s internal tests. Custom Apps are workspace screens a team builds for review. 1:00 p.m. Eastern is the overview’s schema publish clock. The card is the Engine v2 graphic.

PRIMARY here: LangChain’s 25 Sep 2026 overview by Jacob Talbot, schema datePublished 2026-09-25T17:00:00.000Z, plus the four 24 Sep detail posts — Tier A PRIMARY, the company’s own pages. STATUS PRIMARY. The desk label on the chip is CONFIRMED because that is the catalog word for a primary page we can stand on. The Interrupt package, Engine as an in-platform agent, more than 60 million traces on the overview, more than 70 million traces on the Engine v2 post, red teaming from traces and repos, the private beta for existing LangSmith Deployment users, inefficient paths, error rate, latency, and cost, the one-click pull request, SaaS for Plus and Enterprise, self-hosted support coming shortly, bring-your-own-key on the next self-hosted release, user-level memory, the Slack and HTTP memory defaults, Slack file transfer, JSON webhooks, Parallel web search free while Managed Deep Agents is in beta, Trajectories on all plans in the United States, smithtune with Baseten or Fireworks, Public Beta, the internal 96.0 and 53.7 percent tables, and Custom Apps are those pages’. NOT claimed: that this desk attended Interrupt, that hundreds was counted, that 60 million and 70 million are the same count, that Private Beta was printed on the overview, that self-hosted Engine v2 shipped today, that a dollar price was copied, that the August twice-as-good line is a new score in this package, that the fine-tuning tables were rerun, that Parallel stays free after beta, that Trajectories are on every plan outside the United States, a login, a stock tip, or investment advice. The card is the Engine v2 graphic. Distinct from the already-filed docker-cloud-sandboxes, kontext-agent-runtime, autonomyai-discover-mode, and dataiku-agent-management.