Proposed Open Source AI Week 2026 community workshop

Open-Source AI Software Factory

Build a small, inspectable software factory where a GitHub issue becomes a reviewable, evidence-backed pull request. Apache Airflow owns the durable lifecycle. Crabbox supplies disposable remote execution. Humans keep approval and merge authority.

Monday, October 19, 2026 · 2:00–4:00 PM Pacific · San Francisco
Exact venue to be announced · Free · Hands-on workshop + live demo · Target capacity: 40 builders

The systems problem after “the agent can code”

Generating a patch is only one step. A reliable AI engineering system also needs bounded execution, durable identity, retries, human gates, reproducible verification, credential boundaries, retained evidence, and one explicit promotion authority. This workshop builds that control path in the open rather than hiding it behind a product demo.

issue → plan → approval → airflow → crabbox → evidence → pull request

Two-hour build path

  1. 0:00–0:15 · Factory contract. Separate durable authority from disposable compute and define what a worker may mutate.
  2. 0:15–0:40 · Issue to plan. Run the scripted swfactory path, inspect the generated plan, and bind human approval to evidence.
  3. 0:40–1:10 · Durable orchestration. Trace how Airflow owns mapped work, waits, retries, cancellation, and lifecycle state.
  4. 1:10–1:35 · Disposable execution. Move build/test work to a Crabbox runner, stream output, and retain the exact command evidence.
  5. 1:35–1:55 · Failure and convergence. Walk through stale workers, retries, ambiguous mutations, bounded repair, and deterministic promotion.
  6. 1:55–2:00 · Take-home path. Install the skills, clone the open repositories, and adapt the verification contract to your project.

Open-source stack

ProjectRole in the workshopSource
Airflow Software Factory Open reference implementation for issue → plan → approval → execution → review → delivery. GitHub
Apache Airflow Durable lifecycle scheduler for mapped work, retries, waits, and human approval gates. airflow.apache.org
Crabbox Remote execution control plane for leasing a runner, syncing a worktree, running commands, and collecting evidence. crabbox.sh

The workshop is project-first, not vendor-first. The local swfactory replay requires no model key. Remote providers are optional execution backends; no single provider is required to understand or reproduce the architecture.

Design rules we will test

Who should come

Open-source maintainers, AI/agent engineers, platform and CI engineers, orchestration practitioners, and anyone trying to turn coding agents into a reliable engineering system. Bring a laptop if you want to follow along; observers and first-time contributors are welcome.

Suggested local prerequisites: Git, Python 3.12, and uv. Crabbox is optional for the first path; a remote-execution demo will be shown separately.

Open and inclusive

This is a free community workshop intended to be open to all attendees regardless of race, gender, background, employer, seniority, or prior open-source experience. We especially welcome first-time contributors and people who want to learn the systems layer behind agents. Respectful participation is required.

Host

Yossi Eliaz, PhD — Principal Engineer (AI), working on agent execution, software factories, CI runners, isolated compute, and scientific-computing systems.

Open Source AI Week 2026

This workshop is being submitted as a community event for Open Source AI Week 2026, the Bay Area's decentralized open-source AI celebration running October 16–25, 2026. Approval/listing is pending; this page does not imply endorsement by the event organizers or by the open-source projects referenced above.

RSVP

RSVP on GitHub

React with a thumbs-up or leave a comment with the workflow you want to test. The exact venue and any final setup notes will be posted on the RSVP issue.