If you need to approve a production agent stack now, LangGraph is the easier choice to defend on the available evidence. LangChain’s product documentation identifies LangGraph as an MIT-licensed, free library, documents approval checks and cross-session memory, and describes both self-hosted and paid LangSmith deployment options. Vercel’s eve 1.0 announcement makes an appealing case for built-in multi-tenancy, including customer-specific credentials and approval rules, but those capabilities are currently vendor-reported. The announcement does state a versioning commitment: “eve follows semantic versioning, so breaking changes to stable APIs only happen in major releases,” with experimental and internal APIs excluded. That is announcement text rather than shipping documentation, and it does not include eve’s repository or license file.
That does not establish that LangGraph is the better engineering choice for every team. It changes how you should adopt each one: LangGraph is the citable default when license and deployment terms must survive procurement or security review today; eve belongs behind a verification gate, with a pilot only after its repository and operating semantics check out.
The choice also is not simply “own eve” versus “rent LangGraph.” Both frameworks let you own the agent loop. LangChain additionally offers to run parts of the deployment for you. The consequential decision is which controls you want directly, and which operational obligations your team is prepared to accept with them.
The own-versus-rent line runs through the agent loop
An agent harness is the software surrounding the model: constructing inputs, selecting tools, maintaining memory, executing actions, and deciding when a person must approve the next step. This layer can matter independently of model quality. In the independent Self-Harness study, harness-only changes produced study-reported Pass gains of up to 40.6 percentage points across nine model-benchmark combinations with models held fixed. That result does not predict a gain from either eve or LangGraph. It does show why treating the harness as disposable plumbing can be a mistake.
Approval controls are part of that engineering surface, not just an interface feature. Code as Agent Harness describes a Plan-Execute-Verify pattern in which plans constrain intended changes, execution happens inside sandboxed and permissioned environments, and deterministic checks plus human review decide whether to accept, revise, escalate, or roll back the result. The same paper identifies permission-policy specification, side-effect prediction, secret handling, and sandbox-escape prevention as open problems.
That is the ownership bargain. A framework may provide the breakpoint or approval primitive, but your team still has to decide which actions are safe, which credentials they can use, what evidence a reviewer sees, and what happens when execution fails halfway through.
What eve 1.0 puts on the table
Vercel announced eve 1.0 as generally available on November 17, 2026. Vercel also reports shipping more than 230 releases and merging thousands of pull requests since the June public preview. That activity indicates a rapidly developed project; the release count does not by itself address interface stability, which the announcement covers separately with its semver policy.
All of the following eve capabilities are vendor-reported from that announcement, rather than independently verified:
- Control of the loop: Vercel defines eve’s harness as the loop that calls models, runs tools, and delegates to subagents. Developers can choose the model, add or replace tools, edit instructions, and decide which actions require approval.
- Dynamic per-run definitions: Before each model call, eve can load models, connections, subagents, tools, skills, and instructions from the session and external data. In principle, that gives an application a place to apply tenant-specific behavior, although the announcement does not establish isolation or security properties.
- Tenant-scoped access and policy: Vercel reports that one deployment can connect to each customer’s systems using that customer’s credentials and approval rules. Your application authenticates the user and identifies the tenant, after which memory, connection credentials, and approval policy can be set for that tenant and user.
- Cross-session memory: Vercel reports native memory with developer-selected scope. Its built-in file memory uses in-memory storage during
eve devand a private Vercel Blob store when configured on Vercel; third-party memory providers are also listed. - Workflow persistence: Vercel states: “The open-source Workflow SDK saves progress as the agent runs, so it can pick up from its last completed step after a crash.” Resumable workflow tasks can also “wait for a person, webhook, or timer without keeping compute running, and their runs survive restarts.” Resume-from-last-completed-step is a stated behavior; retry idempotency, durability boundaries, and what happens when a step fails mid-action are not stated.
The tenancy proposition is the most distinctive part of the announcement for a software vendor building a customer-facing agent. It suggests a deployment model where tenant context is supplied to the harness rather than separately reconstructed in every tool call. That is a useful design direction, not proof that tenant isolation has been independently tested.
LangGraph’s advantage is a clearer evidence trail
LangChain’s LangGraph documentation describes low-level primitives for single-agent, multi-agent, and hierarchical control flows, with support for any model provider. It documents human-in-the-loop checks for steering and approving actions, built-in conversation memory across sessions, and token-by-token streaming. These are vendor-documented capabilities too; “documented” means citable in product materials, not independently validated.
The operational surface is also unusually visible. LangChain’s free 55-lesson LangGraph course covers state schemas, breakpoints, editing state and human feedback, dynamic breakpoints, time travel, LangGraph Store, short- and long-term memory, deployment concepts, and double texting. That curriculum does not prove LangGraph is easy. It shows that its control points have names, documented behavior, and teaching material your team can inspect before committing.
The commercial boundary is reasonably clear as well. LangGraph itself is described as MIT-licensed and free. LangSmith Agent Server is available in the Plus and Enterprise plans, with Cloud, Hybrid, and Fully Self-Hosted deployment options. Cloud is fully managed; Hybrid combines a SaaS control plane with a self-hosted data plane, and LangChain says no data leaves your VPC in that configuration. Running the open-source library yourself and buying LangSmith’s deployment machinery are separate decisions.
LangChain’s FAQ also says rival frameworks “fall short for complex tasks bespoke to a company’s needs.” That is competitive positioning, not a measurement, and it should carry no more weight in a selection process than Vercel’s performance headline.
The decision axes, with their evidence status attached
| Decision axis | eve 1.0 | LangGraph | Practical consequence |
|---|---|---|---|
| Model and tool control | Vendor-reported: model choice, tool replacement, instruction editing, and dynamic definitions before model calls, according to Vercel. | Vendor-documented: low-level control flows and any-model-provider support, according to LangChain. The two are not compared on tool-substitution mechanics in either vendor’s material. | Both support owning the loop, but the evidence does not establish equivalent programming models or migration effort. |
| Approval gates | Vendor-reported: action approvals plus per-tenant approval policy. | Vendor-documented: human-in-the-loop checks; LangChain’s pages do not document a per-tenant approval policy. | If customer-specific policy is central, eve’s proposed packaging is attractive but must be tested. LangGraph’s generic approval primitive has the clearer documentation trail. |
| Multi-tenant credentials | Vendor-reported: customer-specific connection credentials selected after your application identifies the tenant. | Not established: LangChain’s product pages do not document an equivalent per-tenant credential mechanism. | Do not interpret missing documentation as proof LangGraph cannot support your design. It means you must design and verify that layer rather than assume feature parity. |
| State and memory | Vendor-reported: scoped cross-session memory; the Workflow SDK is described as saving progress as the agent runs and resuming from the last completed step after a crash, though retry idempotency, durability boundaries, and mid-action failure behavior are not stated. | Vendor-documented: conversation histories across sessions; the course also covers LangGraph Store and memory types. | Conversation memory and resumable execution are different requirements. Specify which one you need before comparing checkboxes. |
| Deployment and data control | Vendor-reported: agent and site can deploy together or separately on Vercel. The announcement does not state general self-hosting terms or data-path details. | Vendor-documented: MIT library plus paid LangSmith deployment tiers, including Hybrid with a self-hosted data plane and a Fully Self-Hosted option. | LangGraph currently offers the clearer own-versus-rent boundary. |
| License and supply chain | Unverified: the announcement calls eve open source without stating license terms. | Vendor-documented: MIT-licensed and free as a library. | A team that must cite license terms now has a concrete reason to prefer LangGraph or wait on eve. |
The table is not a feature-equivalence claim. It shows where the evidence supports a decision and where it only supports a question for the vendor or repository.
The 7.3x speed claim is not an eve versus LangGraph result
Vercel’s headline latency result is vendor-reported and narrow. In its hosted stress test, Vercel reports that median turn latency fell from 3.036 seconds in September to 414 milliseconds for 1.0, with p95 falling from 3.692 seconds to 633 milliseconds. The 100-turn test used one immediate mock-model response and no tool calls per turn, deliberately excluding model-generation time to measure eve and Workflow overhead.
That can be useful for tracking orchestration overhead between eve versions. It cannot tell you how either framework will perform with a real model, external tools, approval pauses, or your persistence backend. It is not a head-to-head benchmark against LangGraph.
There is also an unresolved inconsistency: the same Vercel page elsewhere gives the 1.0 median as 626 milliseconds, while the workload description gives 414 milliseconds. Until Vercel reconciles those figures, neither should become the basis for a capacity plan.
Vercel additionally presents a vendor-curated customer account claiming 25,000 fewer lines of code, halved AI costs, and a 66% runtime improvement. Those are adoption anecdotes selected by the vendor, not independent operational comparisons. I would not migrate a working production harness on the strength of either the latency headline or that account.
The recurring bill is evaluation and control, not just hosting
Owning the harness gives you control over routing, tools, data handling, and upgrade timing. It also leaves your team responsible for demonstrating that those choices still work after each change.
A Meta production study makes that cost concrete. The researchers collected 574 benchmark runs over 52 days from an evolving analytics agent as models, prompts, tools, and surrounding systems changed. This is one study-reported operating history, not a forecast that your team will need the same cadence. It does show why evaluation is a recurring engineering function rather than a one-time launch gate.
Managed deployment can remove some infrastructure work. LangChain, for example, describes its Cloud option as automatically updated and maintenance-free. That does not transfer responsibility for deciding whether a new model, prompt, tool, or approval rule is acceptable for your customers. Nor does Vercel’s vendor-reported improvement in eve observability establish equivalent evaluation or incident-response capabilities; the announcement does not provide enough detail for that comparison.
For multi-tenant agents, the security obligation is sharper. If your application supplies the tenant identity that selects credentials, memory, and approval policy, then tenant identification becomes part of the agent’s security boundary. That is an inference from Vercel’s described architecture, not a reported vulnerability. It means your review should include incorrect tenant mapping, credential revocation, cross-tenant memory access, and failures that occur after an action has begun.
Make repository verification the price of an eve pilot
Before committing to eve, I would require five checks:
- Find the actual eve repository and read its license and dependency terms. The vercel/vercel repository is Vercel’s CLI monorepo, whose README states Apache 2.0. It is not an eve repository and cannot establish eve’s license. Research on license drift in open-source AI also warns that automated tools cannot resolve every inherited upstream conflict, so a repository badge alone is not the end of supply-chain review; note that this paper’s authors direct readers to a newer, superseding study.
- Read the upgrade guide and the versioning policy before you rely on either. The announcement states that “eve follows semantic versioning, so breaking changes to stable APIs only happen in major releases,” and that “Experimental and internal APIs, including experimental fields, can change between major releases.” It also points to a 1.0 upgrade guide covering changes to extension builds, agent APIs, session events, clients, and saved history. Both statements are announcement text, not the tagged 1.0 documentation; confirm which side of the stable-versus-experimental line your design sits on before committing.
- Specify persistence before testing it. Determine whether you need conversation history, durable workflow progress, idempotent retries, or all three. The announcement describes resume-from-last-completed-step after a crash, but it does not define retry idempotency, durability boundaries, or behavior when a failure occurs mid-action. Test process termination and recovery against your actual storage configuration.
- Run tenant-isolation tests with your own threat model. Verify that credentials, memory, tools, and approval policy cannot cross tenant boundaries. A successful demonstration of per-tenant configuration is not the same test as an attempted cross-tenant access.
- Benchmark your real workload against your current LangGraph or custom design. Include model generation, tool calls, approvals, persistence, failures, and evaluation effort. Use the same tasks and data controls on both sides; Vercel’s mock-model test supplies no such comparison.
The OpenAI Agents SDK is deliberately absent from this comparison: no sources for it are included here, and adding a third column would require new research rather than inference.
Until those checks are complete, choose LangGraph when citable licensing and deployment terms are immediate constraints, and consider eve when its tenant-scoped design justifies a bounded pilot and your team can perform the verification work. If operating the deployment is the burden you most need to remove, evaluate paid LangSmith separately from the MIT library decision. None of the available evidence supports treating eve’s GA label or latency headline as proof that it is ready to replace an existing production harness.
Frequently Asked Questions
What is the license for LangGraph?
LangGraph itself is described as MIT-licensed and free.
What deployment options are available for LangSmith Agent Server?
LangSmith Agent Server is available in the Plus and Enterprise plans, with Cloud, Hybrid, and Fully Self-Hosted deployment options.

Join the discussion
Share a useful perspective or ask a question about this article.