Building AI Agents With Eve: Tools, Skills, and Evals
Featuring
Why Eve Matters for Agent Development
Eve gives developers a structured way to build durable backend AI agents. Its filesystem-first approach organizes agent behavior around clear project conventions: instructions, tools, skills, channels, schedules, sandboxes, subagents, and evals.
That structure was the main focus of the session. Instead of wiring every part of an agent manually, Eve provides predictable places to define what the agent can do, how it should behave, and how it interacts with its environment.
Building a Local Agent with Eve
The demo centered on a coaching-style agent built with a local exercise dataset stored as JSON. The goal was to test whether Eve could support a practical workflow without relying on external APIs or paid model providers.
The agent ran against a local Gemma model through Ollama, using an MLX-optimized version suitable for Apple Silicon. This setup made it possible to test tool calling, instruction design, and local inference while keeping the experiment offline.
Tools, Sandboxes, and Agent Behavior
Eve’s sandbox gives an agent access to a controlled workspace where it can read files, write data, and work with project resources. The first step was testing direct access to the sandbox, followed by a more structured implementation using tools for retrieving and filtering exercises.
Designing tools with AI agents in mind is critical. Tools that are too limited or overly strict can actually harm the quality of the output by restricting what the agent can access or infer, while more flexible designs allow the model to better interpret context and deliver more useful results.
Skills, Instructions, and Where Context Belongs
The session also explored the boundary between instructions and skills. Information needed for most interactions is usually better placed in the main instructions, while skills are more useful for domain-specific context that only applies to certain requests.
An exercise recommendation skill was added and tested to see when Eve would load it.
Running, Testing, and Evaluating the Agent
The agent was also run as a server and accessed over HTTP, showing how Eve can expose agents beyond a local terminal workflow. Evals were then used to check whether the agent called the expected tools and loaded the expected skill.
The main takeaway is that Eve does not remove the need for thoughtful prompt design, tool design, safety decisions, or testing. What it does provide is a clear structure for those concerns. For teams building AI workflows, that structure can make agents easier to inspect, extend, and improve.
Watch the full recording to see Eve used to build, run, debug, and evaluate an AI agent step by step.
