LangSmith and Onyx better structure AI agent governance
LangChain and Genpact both presented, on November 16, 2026, governance components for AI agents. One links traces, evaluation, and ownership in LangSmith; the other anchors Onyx in business processes with control and audit.
Two separate announcements, same obsession: putting AI agents back under control. On November 16, 2026, LangChain detailed the reconstruction of LangSmith around persistent recording per agent, while Genpact made Genpact Onyx available to build, run, and govern agents in enterprises. In both cases, the challenge goes beyond simple monitoring: it is about linking what an agent does, who is responsible, and the context in which it operates. For technical teams as well as business management, this change matters because it reduces uncertainty when balancing cost, performance, and responsibility.
Un agent, une identité, des traces qui ne se perdent plus
LangSmith, LangChain’s platform, now associates each AI agent with a persistent record including a named owner, immutable unique identifier, and agent type. Until now, traces mostly arrived by project, without a lasting link to the agent that produced them. From now on, local, development, preproduction, and production environments are linked to the same agent, simplifying end-to-end tracking.
The benefit is quite concrete for engineering teams: less back-and-forth between tracing, evaluation, and deployment tools. A single record provides access to traces, evaluations, Agent Server, and incidents reported by Engine with one click. LangChain also presents this foundation as a base to respond faster to questions of ownership, cost, and performance. In short, fewer hand-crafted spreadsheets, more common system.
Des tableaux de bord alignés sur le type d’agent
The agent type is not just a label. At LangSmith, it determines what is displayed on the dashboard, which data connections are supported, and how the agent is evaluated. Two types are currently available: generalist and coding; a voice type is announced as coming soon. LangChain indicates that this future category will need to consider constraints such as low latency and audio quality, in addition to resolving user requests.
A generalist agent shows its description, recent activity, frequently used models and tools, as well as failures detected by Engine. A coding agent highlights merged pull requests, token costs, and categories of work done, with the possibility to link a GitHub repository to track downstream effects. The idea is simple: compare similar agents on the same basis, rather than mixing apples, pears, and some language models along the way.
Gouverner sans enfermer le runtime
LangSmith can also record agents running on external platforms compatible with the OpenTelemetry Protocol, in addition to agents built internally. This openness allows measuring heterogeneous systems without imposing a single runtime across the enterprise. Cost and quality metrics are then consolidated across multiple environments, while the named owner and detailed traces serve as a starting point when costs rise or performance drops.
The platform remains modular: an agent can keep its runtime and use tracing, evaluation, datasets, annotation, and Engine functions as long as it sends its traces to LangSmith. LangChain specifies that agent recording and the LLM gateway add visibility and controls at the organizational level. The broader part of management, with consolidated views at the enterprise scale and identity management via identity provider, is still under development [to be verified].
Avec Onyx, Genpact mise sur le contexte métier
Genpact, for its part, announced the availability of Genpact Onyx on November 16, 2026, after a presentation on November 12 during its third Agentic Operations Day in New York. The platform is used to build, run, and govern AI agents for enterprise operations. Its engine, Process Intelligence Core, injects into agents the context from workflows, internal policies, exceptions, and human decisions. The starting point is therefore not a fixed process diagram, but the real way work is done in the company.
Genpact says it uses Onyx for its own agentic products, for custom client systems, and for solutions provided by its partners. Priya Vijayarajendran, chief product and platform officer, summarizes the approach as follows: “Enterprises get value from AI when they turn operating knowledge into trusted execution, not simply by adding more agents.” The model relies on autonomy acquired progressively: the agent first proposes, the human validates, then responsibility can increase as results are observed.
Quand l’autonomie se mérite, les audits respirent un peu mieux
In Onyx, every agent action is controlled, traceable, and associated with the reason for the decision. Genpact mentions mechanisms for masking personal data, applying policies, and audit logging according to the chosen configuration. The platform also supports responsible controls on bias and transparency. For compliance and security teams, this is obviously the kind of detail that avoids Friday night cold sweats.
Compatible models, frameworks, and cloud infrastructures can be integrated according to technical requirements. Genpact notably cites AWS, ServiceNow, Microsoft, and Databricks among the technologies likely to be part of the setup. An agent can also be exposed in third-party applications via a context interface, with information adapted to user roles. The tracked indicators can then be linked to existing KPIs, opening the door to outcome-based business models.
Les comptes fournisseurs comme terrain d’essai
Genpact first launches Onyx through AP Suite, an offering for accounts payable that breaks down the work into four modules. AP Capture handles ingestion, classification, and enrichment of invoices; AP Advance manages three-way matching, exceptions, and off-order invoices; AP Trace searches for duplicates and anomalies; AP Assist understands intent and automates supplier responses in multiple languages. The whole aims for more automated processing, while keeping escalations to humans when needed.
Genpact presents its results as projections: up to 80% processing without intervention, over 95% duplicate prevention, and up to 90% automatic responses to supplier requests. The product is available in SaaS mode, without obligation to use Genpact’s managed services; integration, consulting, and support remain separate. Customer data, the company specifies, remain in the environment defined by the deployment model and are not used to train models for other clients.
By the numbers
- November 16, 2026 — public availability date of LangSmith and Onyx.
- November 12, 2026 — initial announcement of Onyx in New York.
- 3 — third Agentic Operations Day of Genpact.
- 80% — targeted processing without intervention for AP Suite.
- 95% — duplicate prevention announced by Genpact.