Vercel’s eve 1.0 cuts workflow overhead for long-running agents
By AgentRiot Editorial
Vercel estimates a 76% drop in Workflows cost for one 50-turn session. An agent paused for approval uses no workflow compute while it waits, but its saved state and other services still matter to the bill.

A Slack agent can draft an email on Friday, ask permission to send it, and pick up on Monday when someone clicks Approve. The turn stays open, but the workflow doesn’t need a running function just to wait. That behavior was already part of eve’s June preview (opens in new tab). With the November 17 general availability announcement (opens in new tab), Vercel’s bigger change is how much work the framework does to save progress while the agent is active.
The eve 1.0.0 release (opens in new tab) was tagged and published on November 15, ahead of the announcement. The npm package (opens in new tab) requires Node.js 24 or later and is Apache 2.0 licensed. eve is the framework that calls models, runs tools and delegates to subagents. The open-source Workflow SDK supplies durable execution; Vercel Workflows is the managed backend eve uses on Vercel. Teams can also self-host with a local or Postgres backend (opens in new tab), with different upgrade behavior and infrastructure costs.
The savings come from writing less during a turn
Vercel estimates (opens in new tab) that a 50-turn session with its Evo Docs agent drops from 5.4¢ to 1.3¢ in Vercel Workflows costs, about 76%. The estimate assumes roughly two model steps per turn and 30 days of retention. That’s one workload, not a price cut or a prediction for the whole agent bill. Workflows meters events, data written and data retained (opens in new tab); Functions and Queues have their own rates, and model calls and sandboxes (opens in new tab) are separate costs. The retention window for finished runs (opens in new tab) also varies by plan.
On the default runtime, eve 1.0’s session steps (opens in new tab) let Workflow SDK 5.5 reference previously saved state and persist changes instead of repeatedly writing the growing conversation. In a separate maintainer test of that change (opens in new tab), a hosted run with a mock model had stored 1.15 MB by turn 99, versus 2.64 MB with the SDK upgrade alone. The tagged code still uses structural deltas under experimental QuickJS, not the same by-reference path. And Vercel’s launch notes (opens in new tab) say image checkpoints can keep a reference instead of copying the image bytes again. The files still exist, and large payloads can still require extra writes.
Vercel also reported a median turn time of 414 ms, down from 3.036 seconds (opens in new tab), with p95 down from 3.692 seconds to 633 ms. The hosted test sent 100 sequential messages and used an immediate mock response and no tools. It measures framework and orchestration overhead, not how long a real model takes to answer or how quickly a sandbox restores. On supported backends, eve starts new runs on Workflow spec 9 (opens in new tab) by default. Each run gets a single orchestrator, so background steps don’t have to trigger a separate replay of the workflow. Existing runs keep the spec they started with.
What wakes an agent up
In the email example, the send tool needs an explicit approval policy (opens in new tab). Slack can show Approve and Deny buttons (opens in new tab), and eve records a pending interaction. The gated send hasn’t started, though tools that don’t need approval can keep working. When an eligible person answers, eve continues with the same turn ID. Denial keeps the send from running; cancelling the turn invalidates the pending approval. A new message from the requester that changes the request can also withdraw an unanswered approval.
A workflow tool (opens in new tab) can also suspend on a timer or an incoming webhook instead of polling. Vercel doesn’t invoke functions for the suspended run (opens in new tab), but opening and resolving the wait generates events that count toward usage, and the run keeps its stored data while it waits (opens in new tab). That says nothing about separate work still running. A sandbox has its own lifecycle (opens in new tab): stopping it can preserve files for the next access, not the live process. The 1.0 release (opens in new tab) no longer wakes a stopped sandbox just because a turn starts. The first operation that needs it pays the restore latency instead.
Saved progress also isn’t a guarantee that an email goes out exactly once. Completed workflow steps can reuse recorded results (opens in new tab), but an interrupted step can run again. Use an idempotency key the mail provider honors, so a retry doesn’t send a second copy if the first reached the provider before the result was saved.
Test an upgrade with something still waiting
On Vercel, in-flight work stays on its original deployment (opens in new tab); an eligible idle session can move to the newer deployment on its next delivery. Don’t delete the old deployment while it still has work to finish. The 1.0 upgrade guide (opens in new tab) calls out two pre-1.0 callbacks the new production deployment won’t accept: replies from remote agents and sign-ins requested inside workflow steps. Local and Postgres deployments can resume runs after a restart on the same eve version with intact storage, but changing eve versions can strand those sessions (opens in new tab).
New projects can start with npx eve@latest init my-agent; existing projects should follow the upgrade guide (opens in new tab) before rebuilding and deploying. In a staging pilot, leave an approval pending across a restart or deploy, then check that answering it resumes the intended turn without duplicating the action. eve 1.0’s OpenTelemetry traces (opens in new tab) identify time spent waiting on approvals, questions, sign-ins and budget prompts. Separate that time from execution, then compare workflow bytes written, model usage and sandbox runtime. A faster turn and a smaller workflow record are measurable wins even if the model bill barely moves.

