Skip to content
sarmakska edited this page May 3, 2026 · 4 revisions

agent-orchestrator

Multi-agent workflows with deterministic replay, durable state, and tool budgets.

Built by Sarma Linux. MIT licence.


What this is

Most agent frameworks are demos with delusions of grandeur. They fall over the moment a tool times out or a model hallucinates a parameter. This orchestrator is built around the assumption that everything will fail, repeatedly, and that you need to debug what your agents did six hours after the fact.

Workflows are typed graphs. State is durable in Postgres. Every step writes a trace event that can be replayed deterministically. Agents have explicit tool, token, and wall-clock budgets, all enforced. Supports supervisor, swarm, and pipeline patterns. Ships with a web inspector to walk runs step by step.

Who this is for

  • Teams shipping multi-step agent pipelines that must survive process crashes.
  • Engineers who have been burned by black-box orchestrators at 2am.
  • Anyone who needs audit trails and deterministic replay for LLM workflows.

Why another orchestrator

The orchestrators that survive production have four properties:

  1. Durable state. Process crashes, resumes from the last checkpoint.
  2. Deterministic replay. Given the trace, reproduce the run exactly.
  3. Hard budgets. Token, tool-call, and wall-clock limits that actually halt execution.
  4. Visible execution. Every step recorded, queryable, replayable from any point.

This has all four. Most don't.

Stack

Node 22, TypeScript 5.6, Fastify, Postgres 17 + Drizzle ORM, Redis BullMQ, Next.js Inspector UI.


Wiki pages

  • Architecture — component diagram, database schema, why each piece
  • Quick-Start — clone, configure, first workflow run, replay
  • Graph-DSL — node options, agent kinds, conditional edges, budgets
  • Roadmap — what is shipped and what is next

Repository

github.com/sarmakska/agent-orchestrator

Clone this wiki locally