Examples

Agents built with Theodosia

Each is a standalone repository you can clone and run, a real agent rather than a snippet.

Repo	What it is
Leavitt	On-call AI diagnostician. Reads metrics (Prometheus), logs (Loki), client load (k6), and feature-flag state through MCP; uses Theodosia’s classified upstream responses so one bad source cannot poison the diagnosis; degrades or declines under chaos instead of guessing. Read-only by construction (the FSM has no write edge). Won the Crusoe track at the DevNetwork AI+ML Hackathon 2026.
Phoebe	SRE incident-investigation FSM. The agent keeps the full Grafana toolset; the FSM gates the procedure (triage, diagnose, verify, conclude) and the audit trail, not the tools. Ships a Harbor agent for Grafana’s o11y-bench.
triage-agent	Support triage: investigate before you decide, enforced by the state-machine graph.
deploy-gate-agent	A change/deploy gate: ordered gates, a health gate, an audit trail, and a call out to a filesystem MCP server through upstream.
coffee-agent	The toy: a coffee-order state machine an LLM drives one enforced step at a time. The smallest interesting graph.

In-repo examples

The examples/ directory in the Theodosia repo ships self-contained FSMs covering the surfaces in What works through mount(): pure-FSM, typed state, lifecycle hooks, persistence, real shellouts, LLM-in-the-graph, SKILL-to-FSM conversions, upstream, and multi-graph. Each runs with uv run python examples/<file>.py.

Start from Authoring a graph to build your own.