From opaque agents to observable AI infrastructure
Open-source agent frameworks and AI agent telemetry tools are emerging as a way to make complex, multi-step AI systems observable, testable, and cost-aware for developers and security teams. Instead of treating agents as black boxes hidden in a SaaS UI, these projects shift control back to engineering teams by exposing how agents are configured, what they do, and which resources they consume. This new stack blends agent-runtime frameworks like Vercel Eve with observability layers such as Agent Beacon, giving organizations a unified view across local development, CI pipelines, and cloud-hosted agents. In practice, that means developers can compose agents as code, trace their decisions through OpenTelemetry, and connect those traces to existing dashboards and SIEM tools. The result is a clearer path to addressing shadow AI, enforcing internal policies, and keeping AI-related compute and tooling costs under closer scrutiny.
Vercel Eve: treating agents as directories, not black boxes
The Vercel Eve framework introduces a file-system centric approach where “an agent is a directory” containing everything it needs to run in production. One file sets the model and connects to Vercel’s AI Gateway for provider fallbacks, another holds the system prompt in Markdown, and each tool is a TypeScript file whose filename becomes the tool name without extra registration. Skill.md files and MCP servers link agents to external tools, while Eve compiles the directory into a running agent backed by durable workflows via Vercel’s Workflow SDK. Every conversation runs as a checkpointed workflow that can pause, survive crashes, and resume. On the security side, each agent executes in its own sandbox and tools can require human approval. To improve AI agent telemetry, every run emits an OpenTelemetry trace that appears in a new Agent Runs view and can be exported to observability platforms like Datadog and Honeycomb.
Developer workflows and cost-aware deployment with Eve
Vercel positions Eve as “Next.js for agents,” extending familiar web-app workflows to AI development. Developers can start agents locally with a single command and talk to them through a terminal interface, then deploy with the same vercel deploy command used for other projects. Sessions that begin on one version of an agent finish on that version, even if a new release ships mid-run, which reduces the risk of breaking long-running tasks. According to Vercel, agents already trigger around 29 percent of deployments on its platform, up from less than 3 percent a year ago, and the company expects that share to reach half. Internally, Vercel runs more than 100 Eve agents, including a data-analysis assistant used tens of thousands of times per month in Slack and a routing agent that forwards questions to specialized peers. This kind of usage illustrates how structured, observable agents can become core infrastructure rather than experimental tools.
Agent Beacon: an open-source telemetry layer for coding agents
Agent Beacon tackles AI agent telemetry from the other side: it instruments what local and cloud coding agents actually do on developer machines, CI jobs, and cloud runtimes. The tool discovers supported runtimes such as Claude Code and Codex CLI and configures them to export OpenTelemetry data to a localhost collector. For Cursor, Beacon installs hooks that emit events for sessions, prompt submissions, tool use, command execution, MCP-like activity, approval decisions, and file edits, where Cursor exposes those payloads. A bundled collector turns both OpenTelemetry logs and hook events into a normalized JSON event log that powers a local dashboard, and teams can forward this data into customer-managed SIEMs. Written mostly in Go and released under the MIT license, Beacon runs locally without an Asymptote account and supports configurable retention modes that can drop prompt text, redact sensitive content, or keep specific fields under size limits so organizations can balance visibility with developer privacy.
Toward governance, shadow AI control, and cost transparency
Beacon’s creators see AI agent telemetry as a foundation for governance rather than an end in itself. Asymptote Labs CEO Justin D’Souza argues that “the precondition for governance is a visibility layer built on a normalized schema of agent activity across local + CI + cloud agent harnesses,” and describes existing endpoint detection products as missing this agent-level understanding. Their roadmap includes a detection rule standard for suspicious agent behavior, a policy layer to gate agent actions in real time, and streaming-first infrastructure so events can be evaluated quickly across enterprise environments. In parallel, frameworks like the Vercel Eve framework embed Agent Beacon observability ideas into agent runtimes by default, making every run traceable. Together, these open source agent frameworks and observability tools help organizations rein in shadow AI, build trust between developers and security teams, and gain the cost transparency needed when agents drive a growing share of deployment and compute activity.
