Learning Eve: Going Further

You built a first agent in the core series: instructions, a model, tools, skills, a channel. That agent can think and act. But a production agent usually needs to reach outside itself — other services, a controlled workspace, specialists, a clock, observability, and a safe path to ship.

What this series covers#

Seven deep dives, each a real capability you can add to the agent you already have:

1. Connections — bring in tools from MCP and OpenAPI services, with app vs. user auth, Vercel Connect, and per-connection approval.
2. Sandbox — run untrusted shell and file work in an isolated /workspace, seed files, pick a backend, and lock down the network.
3. Subagents — delegate parallel or specialist work to fresh agent copies or declared specialists.
4. Schedules — run the agent on a cron cadence, fire-and-forget or handed off to a channel.
5. Hooks — observe the runtime event stream for audit logs, metrics, and alerting.
6. Evaluations — define scored checks and run them on every deploy and on a schedule.
7. Deploy — take the agent from eve dev to Vercel or your own host.

How this series works#

Each step takes one building block through the same path: What it is → Why you'd care → Before → After → Do it yourself → Gotchas → Recap. It builds on the core series, so instructions.md, agent.ts, tools/, skills/, and channels/ are assumed known.

Start with step one.