JobDB
Iceberg for agentic workflows: a durable workflow runtime for job state, retries, wait points, human orchestration, and audit history.
github/colony-2/jobdbApache licensed · Alpha · Contributors welcome
The Colony2 stack for cellular development.
Two open-source building blocks for durable, agentic workflows:
JobDB, a durable workflow runtime, and c2j,
the CLI that submits and runs jobs against it.
$ c2j submit recipe.yaml
job queued
$ c2j run
step recorded · artifacts saved · ready for next lease
Open-source building blocks
Colony2 keeps job ownership separate from job execution. That split lets large, long-running workflows pause, resume, retry, and leave an audit trail while still running on modest infrastructure.
Iceberg for agentic workflows: a durable workflow runtime for job state, retries, wait points, human orchestration, and audit history.
github/colony-2/jobdb
The job CLI. A job is an execution of a recipe, and c2j
submits, runs, lists, and resumes those jobs against JobDB.
JobDB
JobDB defines the durable API surface for jobs. It keeps the workflow state small, explicit, and inspectable while execution can happen elsewhere.
The reference implementation is designed for pluggable storage backends and a scalable split between job ownership and job execution.
c2j
c2j executes recipes inside a cell. Recipes compose state
machines, sequences, and ops. Ops are language-agnostic commands with
defined input and output schemas.
Each op execution can record a git commit and output artifacts such as rollouts, design docs, test statements, and other durable evidence of work.
$ c2j [command]
Available commands:
cells list allowed dependent cells
list list jobs from the JobDB runtime
ready count ready jobs for a tenant
run run jobs through the runtime
submit submit a new recipe job
How they fit together
The Colony2 stack gives cellular development a durable loop:
Git stores the cell, c2j runs the recipe, and
JobDB records state, retries, waits, handoffs, and audit.
Durable job state, retries, wait points, human orchestration, and audit.
Cell content, cell memory, recipe definition, and ops implementations.
Execute jobs, collect user input, list jobs, submit jobs, and run inside a cell.
Config resolved through CEL templates and artifacts pass from one op to the next. The job can commit results back into the cell when the recipe decides that is the right outcome.
Get started
The work is intentionally open source and componentized. Start with the runtime and CLI, then add reusable ops and recipes as your workflows grow.