Vercel eve 1.0 vs LangGraph: which fits a small business
Vercel's eve 1.0 is generally available, so this month's decision is a scoped pilot
Vercel released eve 1.0 on November 17, 2026, making its open-source framework for production agents generally available after a public preview that began in June. The category decision is whether your next internal agent runs on a framework your team hosts and operates, or on a runtime Vercel operates for you, and for most small B2B teams the right move this month is one scoped pilot rather than a migration of automations that already work.
An agent framework is the layer that calls a model, runs tools and keeps a session's state between steps. Vercel's eve packages that loop with sandboxed code execution, durable sessions on Vercel Workflows, chat channels such as Slack, and tracing. It defines an agent as files: a model file, an instructions file, and folders for tools and subagents, as Vercel's June post described it.
A naming check comes first. LangChain's pricing page does not use the name LangGraph Platform and lists deployment under "Agent Server" (LangChain pricing). We could not confirm from the pages we read that the product was formally renamed, so this piece calls the rival LangGraph and takes every plan fact from LangChain's pricing page.
Scope: this is a desk review of public pages opened on November 17, 2026. We did not run eve or LangGraph for this article. Where a vendor page is silent, the comparison says so instead of filling the gap.
TL;DR: eve 1.0 fits a small team that already deploys on Vercel and wants durable agents with approvals and traces without building that runtime. It does not justify replacing a working Zapier, Make or n8n flow, and it does not remove the need for a named owner, approval rules and a cost check.
Key Takeaways
eve 1.0 is generally available as of November 17, 2026, and Vercel says breaking changes to stable APIs happen only in major releases.
The June public preview and the 1.0 release differ most on latency claims, channel coverage, memory, shared Slack sessions and tracing detail.
Vercel's launch post does not publish a price for eve itself, so the real cost is Vercel's usage meters plus model fees.
LangChain's Developer and Plus plans charge per seat, so its cost rises with headcount, while the Vercel pricing page we read lists no per-seat charge.
A pilot needs one owner, one narrow task, human approval on every external write, and a trace review before wider rollout.
A widely reported Gartner forecast expects many agent projects to be canceled, so a pilot with a stop date is the sensible default.
How we evaluated these options
We scored the two options on seven decisions a small B2B team actually makes, and weighted each one by how often it blocks a purchase or a pilot. The weights are our judgment. They are not vendor data or analyst ratings, and we show them so you can change them to fit your own priorities.
| Criterion | Weight | Why it matters for a small B2B team |
|---|---|---|
| Hosting and lock-in | 20% | Decides whether the agent can move later and whether your engineers already run the platform. |
| Durable execution | 15% | Decides whether an interrupted task resumes instead of repeating external actions. |
| Human approval controls | 15% | Decides who signs off on sends, deletes, refunds and other writes. |
| Observability and traces | 15% | Decides whether someone can review what the agent did without writing custom logging. |
| Published cost | 15% | Decides whether a budget can be set from public numbers before anyone requests a quote. |
| Team access and channels | 10% | Decides whether staff reach the agent in tools they already use, such as Slack. |
| Release maturity | 10% | Decides how much API churn to expect over the next year. |
Read the weights as a buyer's priorities, not as a score. A team with a strict data-residency rule would move hosting and lock-in above everything else, and should re-weight the table before deciding.
What changed from the June preview to the 1.0 release
The June preview is the right baseline because most buyers who looked at eve before November saw that version first. InfoQ dated its report June 26, 2026, and described Vercel's release as a public preview, according to InfoQ. Techzine's June 24, 2026 report placed eve alongside Passport, an identity layer for applications and agents, announced at Vercel's Ship conference, according to Techzine. Passport is a separate product and is not part of the 1.0 release.
| Area | Before (June preview or September beta) | After (1.0, November 17, 2026) |
|---|---|---|
| Release status | Public preview | Generally available; breaking changes to stable APIs only in major releases |
| Median hosted turn latency (100 turns, mock model) | 3.036s on the September 1 beta baseline | 414ms on November 15, 1.0.0 |
| Estimated Workflows cost, 50-turn session | 5.4¢ | 1.3¢, about 76% lower under Vercel's assumptions |
| Native channels | HTTP API, Slack, Discord, Teams, Telegram, Twilio, GitHub, Linear named in the June post | Adds Web Chat and iMessage; Chat SDK adapters add WhatsApp, Messenger, Google Chat and email |
| Multiplayer sessions | Not described in the June post we read | Shared Slack threads with per-person approvals and an ambient mode |
| Memory | Not described in the June post we read | Built-in file memory plus named external memory providers |
| Tracing | OpenTelemetry-based tracing of model calls, tools and sandbox commands, per InfoQ | Traces show waits for people, model choices, approval decisions and calls into MCP servers |
| Dynamic definitions | Not described in the June post we read | Models, connections and subagents load per call, alongside tools, skills and instructions |
The latency and cost rows, 414ms and 1.3¢ at the 1.0 release, come from the launch post, according to Vercel (2026). Vercel measured latency as overhead only, using a mock model and no tool calls, so 414ms says nothing about how fast a real model answers. Its cost estimate assumes about two model steps per turn and 30-day retention.
The launch page is also internally inconsistent. Its chart caption gives a different 1.0 median than its table does, and the post does not reconcile the two. This article uses the table value. Ask Vercel which figure is correct before quoting either one in a budget or a board update.
The change that matters most to a small team is not raw speed. Approvals, channel routing, memory and tracing are now described as parts of one runtime, which means fewer separate pieces for your team to wire together. The trade-off is that those pieces are Vercel services, so the decision depends on how comfortable your engineers are operating that platform.
eve 1.0 against LangGraph on the decisions that matter
| Decision | Vercel eve 1.0 | LangGraph (LangChain) |
|---|---|---|
| Release status | Generally available as of November 17, 2026 | Not verified in this review |
| Pricing model | Plan tiers plus usage meters; Enterprise is quote-based | Per-seat plans plus base trace allowances; Enterprise is quote-based |
| Trace visibility | Agent Runs in Vercel Observability; OpenTelemetry export for self-hosted deployments | Base traces kept for a set period; extended traces are a separate option |
| Durable execution | Sessions run on Vercel Workflows and resume from the last completed step | Not verified in this review |
| Approval controls | Per-tool approval policy; a tool runs without approval unless a policy is set | Not verified in this review |
| Team access | Shared Slack threads with individual approvals and an ambient mode | Per-seat access; collaboration features not verified |
| Hosting | Deploys to Vercel; the June post said other platforms were on the way | Managed Agent Server metered by resource; Enterprise adds self-hosted and hybrid options |
| Release stability | Semantic versioning; breaking changes to stable APIs only in major releases | Not verified in this review |
The Vercel-side cells come from its launch post, its June post and its documentation. The LangGraph-side cells come from LangChain's pricing page. We did not open LangGraph's framework documentation for this piece, so its runtime behavior is marked not verified rather than guessed.
Vercel eve 1.0: best fit, limits and implementation
eve 1.0 fits teams that already deploy on Vercel and want agents that pause for approvals, resume after a crash and show traces without building their own workflow engine. The launch post says the open-source Workflow SDK saves progress as the agent runs, so it can pick up from its last completed step after a crash.
The limits are real. The June post said deployment to other platforms was "on the way," and the pages we read do not publish a self-hosted production guide. The launch post says that teams who host eve themselves can export traces to an OpenTelemetry backend, which is a different thing from a documented production path. InfoQ also reported that some commenters raised portability as an open question. Passport, the identity layer announced alongside eve, requires an Enterprise plan.
For implementation, Vercel's docs list Node.js 24 or later, pnpm, a Vercel account and team with AI Gateway access, and the Vercel CLI. The agent starts with pnpm dlx eve@latest init my-agent, the model setting lives in agent/agent.ts, and access rules live in the channel configuration. The docs also state that a tool call runs without approval unless you set a policy, so approval rules are an explicit setup task rather than a default you can assume.
Choose it if your team has an engineer who will own agent code and your work already centers on Vercel or Slack. Skip it if no one will maintain agent code after launch, or if your stack must avoid a single hosting vendor.
LangGraph: best fit, limits and implementation
LangGraph, sold through LangChain's plans, fits teams that want to design the agent's control flow explicitly and host it on infrastructure they pick. LangChain's pricing page offers a seat-based Plus plan and an Enterprise tier with self-hosted and hybrid options, which matters if keeping the runtime off a single vendor is a requirement.
The limit is evidence. We did not open LangGraph's framework documentation for this piece, so we cannot describe its current APIs, durability behavior or approval features with a citation. InfoQ's June report named LangGraph among the frameworks it compared against eve, but a comparison mention is not evidence of how LangGraph behaves in production.
For implementation, check LangChain's documentation, build the same small agent on both stacks, and compare the trace output and approval steps on one identical task before choosing.
Choose it if your team wants explicit graph control and can run a separate deployment target. Skip it if the goal is one managed runtime with no framework to maintain.
Pricing and usage meters
The launch post does not state a price for eve itself, so the vendor has not published one. The costs a pilot will see are the Vercel services eve uses, including Functions, Workflows, Sandbox, Blob and AI Gateway, plus model fees. The tables below show published plan prices and the meters a pilot would touch.
| Vendor | Plan | Published price | Included or limit |
|---|---|---|---|
| Vercel | Hobby | $0 per month | 50K Workflows events and 1M function invocations per month |
| Vercel | Pro | $20 per month | $20 included credit across resources; usage billed by meter |
| Vercel | Enterprise | Quote-based | Custom limits |
| LangChain | Developer | $0 per seat per month, then pay as you go | 1 seat; up to 5k base traces per month |
| LangChain | Plus | $39 per seat per month, then pay as you go | Unlimited seats; up to 10k base traces per month |
| LangChain | Enterprise | Quote-based | Custom seats and workspaces; self-hosted and hybrid options |
Pricing checked November 17, 2026. Enterprise rows are quote-based because neither vendor lists a public price for them.
| Meter | Hobby included | Pro published rate |
|---|---|---|
| Workflows events | 50K per month | $20 per 1,000,000 events |
| Workflows storage writes | 1 GB | $0.50 per GB |
| Workflows storage retention | Not available | $0.50 per GB-month |
| Fluid Active CPU | 4 hours per month | Starting at $0.128 per hour |
| Function invocations | 1M per month | $0.60 per 1,000,000 |
| Sandbox snapshot storage | 15 GB | $0.08 per GB-month |
On Vercel's Pro tier, Workflows events cost $20 per 1,000,000 events and Fluid Active CPU starts at $0.128 per hour, according to Vercel (2026). The pricing page also says the Pro plan includes a $20 credit across resources, which is a separate allowance from the meter rates above.
The LangChain side is simpler to read. Plus plan: $39 per seat monthly according to LangChain (2026), with up to 10k base traces a month before pay-as-you-go rates apply.
A worked example for a four-person pilot
Take a hypothetical four-person operations team that runs one internal agent in Slack, with its model setting in agent/agent.ts. On the Plus plan from LangChain, four seats at $39 each come to 4 × $39 = $156 per month before usage, and the 10k base trace allowance applies first, according to LangChain (2026). On the Vercel side, the launch post estimates the Workflows cost of a 50-turn session falls from 5.4¢ to 1.3¢. If the team ran 200 such sessions a month, that is $10.80 before and $2.60 after, a difference of $8.20 for Workflows alone. The math excludes model tokens, sandbox time and Blob storage, and it assumes Vercel's two model steps per turn and 30-day retention. Treat the 200 sessions as a placeholder to replace with your own count.
Who this is for
Who this is for: a founder, operations lead or engineering lead at a small B2B software company that already deploys on Vercel, or that has an engineer who can own an agent service after launch. The fit is strongest for one repeatable internal task with a named reviewer, where the agent reads from systems with APIs and writes back only after a person approves.
Red flags: no one will own the agent after launch; the first task sends customer messages or moves money with no approval step; or the systems involved have no API or export path.
The market case for caution is real, and the pilot should be designed around it. The Star reported a Gartner forecast that expects cancellations driven by escalating costs and unclear business value, with the deadline set at the end of 2027. Forecast: more than 40% of agentic AI projects canceled according to The Star (2025).
Build, buy or stitch it together
The real alternative for many small teams is a stitched workflow in Zapier, Make or n8n, or an in-house script. Those tools can keep run histories, retries, error branches and audit evidence when someone configures them. In every case, the buyer designs and owns observability, idempotency, escalation paths, access controls and ongoing maintenance. eve moves more of the agent logic into one runtime, but it also means your team versions agent code and reviews it like any other service.
A proposed US Tech Automations workflow for refund review would start when a support ticket in your helpdesk is tagged for refund review. It would read the order and account record from your billing system through its API, draft a recommendation citing the refund policy clause it applied, and write that draft back to the ticket as an internal note. Prerequisites are API or export access to the helpdesk and billing system, read-only scope on the billing data, and a written refund policy the workflow can cite. The human review point is a staff member who approves or rejects every refund before any payment action runs.
A second proposed workflow would start when a new customer request appears in a shared Slack channel. It would create a task in your project tracker through its API, attach the thread link, and post a status reply in the thread when the task changes state. Prerequisites are a Slack app with channel access, a tracker token scoped to one project, and a written rule for which requests qualify. The human review point is the project owner, who sets priority, so the workflow never assigns priority on its own.
What US Tech Automations could configure differently from a stitched Zap or flow is one place for retry limits, duplicate-prevention keys, escalation routes and audit retention, instead of those rules living on several separate canvases. That is a design description. We have no customer results for these configurations to report.
When NOT to use US Tech Automations: skip it if a single built-in feature in your existing tools handles the task, if you have no API or export access to the systems involved, or if the work is a one-off report you can run by hand. In those cases a simpler existing tool wins on setup time and cost.
Decision checklist for this month
- Name one owner who can change the agent's instructions, tools and approval policy.
- Pick one internal task with a reviewer who checks output before any external write.
- Confirm which Vercel plan and meters the pilot needs, using the pricing tables above.
- Set approval on every tool that sends, deletes, pays or publishes before the first live run.
- Decide who reads the traces, how long they are kept, and whether you need an export.
- Set a stop date: if the pilot does not beat the current workflow within 30 days, stop.
Signal vs Speculation
Signal, stated by a vendor or a named outlet
Vercel says eve 1.0 is generally available and that breaking changes to stable APIs happen only in major releases.
Vercel does not publish a price for eve itself.
Vercel's latency figures measure overhead with a mock model and no tool calls, not real model response time.
The launch page's chart caption and its table give different 1.0 median latencies.
The customer examples, including Rillet and Clay, are Vercel's own claims, and we did not verify them.
Passport requires an Enterprise plan.
Real agentic AI vendors: about 130 of thousands according to OODA Loop (2025), which reports Gartner's estimate.
Speculation, our read
The 1.0 label lowers buyer risk over the next year only if major-version changes stay rare; one release cannot confirm that.
Shared Slack sessions help a team that already runs agents in Slack more than a team that uses an agent alone.
Many small teams can run one agent per task without a separate orchestration layer, while teams running several agents across tools will need one.
Portability will matter more than the launch post suggests, because the durable production path and sandboxes run on Vercel services.
Frequently asked questions
Does eve 1.0 replace Zapier, Make or n8n?
No, eve is a code framework for building agents, while Zapier, Make and n8n are workflow tools you configure, so keep working automations in place and pilot one agent beside them.
Does Vercel publish a price for eve 1.0?
No, the launch post does not state a price for eve itself, so cost comes from the Vercel services it uses plus model fees.
What changed between the June preview and 1.0?
The release moved eve from public preview to generally available and added memory, shared Slack sessions, dynamic definitions and richer tracing, and the comparison table above lists each change against the earlier version.
Can a small team run eve outside Vercel?
Possibly, but the pages we read do not publish a self-hosted production guide, so plan for Vercel hosting unless your engineers verify a self-hosted path first.
Should a small team act on eve 1.0 this month?
Act only if you can name one owner and one narrow internal task with a reviewer; otherwise, wait until those exist before starting the pilot.
Next steps before the month closes
Pick the one task, assign the owner, put approval on every external write, and read the traces for two weeks before widening access. For a parallel launch explainer, see Agent Plugins 1.0 explained. For another launch written in this format, see Tiger Neo 5.0 explained. For a separate SaaS model comparison, see GPT-6 Astra vs Claude Fable 5.1 for SaaS. If you want to see how US Tech Automations configures this for a specific stack, see how US Tech Automations configures this.
About the Author

Helping businesses leverage automation for operational efficiency.
![Trim Chiropractic Email Sequences 2027 [Decision Guide]](https://images.pexels.com/photos/6129037/pexels-photo-6129037.jpeg?auto=compress&cs=tinysrgb&fit=crop&h=630&w=1200)