Open SourceRuby / SQLite / RubyLLM

Tamoz: durable agent workflows for Ruby

Agent FrameworkDurabilityEvaluation
View Repository
ProblemAgent runtimes need durable execution, explicit review boundaries, and evidence before they can be trusted with real work.
ShapeRuby monorepo with core protocols, graph execution, SQLite persistence, agent runtime, and eval gems.
StatusIn active development; core, graph, SQLite, agent, and evaluation foundations are underway.

Tamoz is a Ruby-native durable agent framework for checkpointed, interruptible, observable, and evaluation-governed AI workflows. It is structured as a monorepo of independently publishable gems rather than one large runtime.

The current build is risk-first. Evaluation contracts come before broad agent behavior. Deterministic graph execution comes before persistence. Persistence comes before larger model-driven workflows. The result is slower than a demo framework, but the boundaries are easier to inspect.

Why I built it

Most agent systems fail at the edges: repeated side effects, ambiguous recovery, hidden state, unreviewed edits, weak verification, and no durable evidence trail. Tamoz treats those as first-order product requirements instead of cleanup work after the agent can already act.

What to inspect

  • Core runtime values for context isolation, cancellation, streams, and state codecs.
  • Deterministic graph execution with checkpoints, interrupts, reducers, and subgraph support.
  • SQLite durability for checkpoints, effect journals, leases, backups, and boundary audits.
  • Tamoz Agent's reviewed plan-before-action loop with workspace-confined tools.
  • Evaluation artifacts, scorecards, schemas, and conformance fixtures that describe release evidence.

Why it belongs here

Tamoz reflects the same engineering thesis behind my other agent infrastructure work: autonomy needs contracts, recovery paths, observability, and evaluation gates. The project is still in progress, but the architectural direction is already visible in the code and docs.