Skip to main content
Back to timeline
Vercel NewsSource publication:

eve 1.0 goes generally available: the same 100-turn stress test drops from 3.036 s to 414 ms per turn

Research institutions:Vercel

Synopsis

Vercel has moved its open-source agent framework eve from public preview to a 1.0 general release, adding security approvals, shared multiplayer sessions, cross-session memory and OpenTelemetry tracing. In a hosted stress test on Vercel Workflows, the same 100-turn workload fell from a median of 3.036 s per turn in early September to 414 ms; each turn sends one message, gets an immediate mock-model response and makes no tool calls, and timing excludes model generation. The framework hands users the loop that calls models, runs tools and delegates to subagents, with a resumable Workflow SDK built in.

Source-provided article image: Introducing eve 1.0
Vercel

Interpretation

The open-source agent framework eve leaves public preview with a 1.0 release that leaves model choice, adding or replacing tools, editing instructions and which actions need human approval to the user. After the June public preview, the team worked with early adopters and shipped more than 230 releases and merged thousands of pull requests to bring eve to general availability. Version 1.0 makes eve faster, expands integrations and improves observability, and adds security controls that let users decide which actions need approval and what data gets collected.

At the center is a harness the user owns: the loop that calls models, runs tools and delegates work to subagents. The open-source Workflow SDK saves progress as the agent runs so it can resume from its last completed step after a crash, and multiplayer sessions and agent-to-agent calls are built in. Managed hosting runs on Vercel, where Vercel Workflows handles durable execution and Vercel Sandbox isolates agent-generated code; Rillet's software factory on eve turns customer requests into specs, tests changes and opens PRs for human review, and tripled its shipping rate in three months.

Version 1.0 rebuilds tracing around OpenTelemetry so an agent's work shows up in the tracing tools users already run. Traces now show when an agent is waiting for a person, including tool approvals, questions, budget prompts and sign-ins, and record model selections and approval decisions alongside model usage and cost. eve dev records traces locally by default, and self-hosted deployments can export them to an OpenTelemetry backend including one they run themselves; private conversations in production send only metadata by default, without prompts or tool results.

Perspective

The demonstrated scope is the 1.0 general release of the open-source framework and its latency, cost and tracing behavior under Vercel managed hosting; the latency figure comes from a 100-turn stress test with one message per turn, an immediate mock-model response and no tool calls, and timing excludes model generation.

The roughly 76% cost drop is an estimate assuming about two model steps per turn and 30-day retention, using Vercel Workflows pricing; on Vercel, inputs above the payload size limit still need extra storage writes, and images and files can trigger those writes.