Devy

The six-agent detour

My first design had six agents: infrastructure, CI/CD, performance, observability, connectivity, and security. Each had its own MCP server. A central hub coordinated them with a notification protocol I'd spent a lot of time thinking through.

In practice, they burned tokens explaining context to each other and second-guessed one another. When the model made a bad call, several agents often made the same bad call. The extra machinery made the source of an error harder to find.

I ended up with one capable agent behind a central service. The hub disappeared. MCP became a way to expose tools, which was all I needed it to be.

One service, several ways in

The service owns the reasoning loop, tools, memory, and tracing. A web chat, a native Go command called ask, and an HTTP endpoint all use the same API. I wrote the harness directly: assemble context, call the model, dispatch its tools, repeat.

A tools-router lets the agent discover relevant tools through find_tools instead of loading the entire catalog into every conversation. Runbooks and postmortems go into a hybrid vector and full-text index, so the agent can retrieve and cite the documentation alongside live observations.

An investigation you can follow

The most useful capability emerged while building it. With logs, container state, and documentation in one context, the agent could assemble an incident timeline across all three.

The repo includes a deliberately crash-looping container to try this on. Devy checks the logs and host state, finds the relevant runbook and past incidents, and ranks the likely causes. In that example, it distinguishes an out-of-memory symptom from the exhausted connection pool behind it. You can inspect the evidence rather than take its word for it.

What it can touch

Hosts are exposed through declarative, profile-gated tools. There is no shell access. The agent can investigate and recommend; consequential production actions need a person.

That boundary matters when you're asking someone to connect an AI to their infrastructure. It also keeps the system easier to reason about when something goes wrong.

Try it on your own infrastructure

Devy is Apache-2.0 licensed and runs with docker compose up. The repository includes the demo, tests, and a JOURNEY document covering the designs I tried and the reasons I changed them.

The trading-platform details stay private. The framework is there for anyone who wants a starting point for their own operations assistant.