The problem
"The AI got good. The tools for long-running work did not." POSTER
Agent work now runs for hours, days, weeks. A process that long will crash, take a wrong turn, or need a human's sign-off — and most tooling answers all three with "start over." POSTER AI
Crashes
The laptop dies at transform 23 of 41 — "the process is gone; the work isn't" is the standard nobody meets. POSTER
Wrong turns
Six files migrated with a deprecated codemod. "Retrying won't help; the bad edits are upstream, already in the code." POSTER
Lost work
Without checkpoints, every resume is a restart from zero. POSTER
What is it
"Durable multi-agent workflows from a single prompt." Open source, MIT. POSTER
Their approach in three steps: write a prompt → generate a durable workflow → monitor it. Durability, in their words: "It's why you can order an Uber and put your phone in your pocket. Same for agent work: every step checkpointed." The contract on the poster: "a completed task is never re-executed" — resume skips the 23 done transforms and continues to 41 of 41. POSTER
The live site says it plainly: "Agent workflows with time travel … describe the workflow in plain English and your coding agent builds it — no files written by hand. Then watch every step stream live, and rewind, fork, or replay any run from any point." WEB
What it does
The poster's key features, verbatim.
Time travel
"Rewind, fork, replay · git for work." POSTER
Zero config
"Plain English in" — no workflow files written by hand. POSTER
Any model, any harness
"Switch freely" — Claude Code, Codex, OpenClaw, Hermes, "+ 9 more CLI harnesses · kimi, opencode, amp, vibe, forge, …". POSTER
Self-improving
"Optimizes its own prompts." POSTER
100 example workflows
"Every orchestration pattern out there, an example of anything we saw twice" — in the repo's examples/. POSTER
How it works
Describe the job; the agent authors the workflow; every step lands in a durable store the process can die around.
Describe
The time-travel demo on the poster: a zod v3 → v4 migration across 41 files hits a wrong turn; smithers replay … --frame 35 --restore-vcs rewinds to the frame before the bad transforms — code and run state move together. "One crash · one wrong turn · zero lost work." POSTER
Workflows are authored on top of React — "a mature framework agents are already trained to be really good at using" PITCH. In the default runtime "every transition, attempt, and approval is a SQLite row" WEB; in Aomi's production deployment the smithers store runs on Supabase Postgres as "the single durable truth" POSTER.
The numbers
The poster's figures, re-checked live against npm and the GitHub API on 2026-07-22.
Who's building it
One creator, full time, with an open-source contributor bench behind him. REGISTRY
Will Cory
"Hi, I'm Will, the creator of Smithers" — his own opener at the cohort's demo-day rehearsal. Top contributor on the repo by three orders of magnitude. The cohort profile lists prior work: Tevm, an EVM toolkit backed by an Ethereum Foundation grant, and Guillotine.
The lineage
Applied to the cohort as JJHub, a collaborative cloud coding platform; pivoted in week 2 to focus solely on Smithers. jjhub.tech is still live as a legacy product of the same org — "collaborative cloud coding for teams and agents."
Who builds on it
"The OSS community builds on it: multiple third-party UIs, a smithers workflow app store, Kubernetes integrations." POSTER
Aomi /build — Exhibit A
The poster's production case: "one prompt → shipped project." Under Aomi's /build page, every build runs as a durable smithers run — supervisor, heartbeat, replay from the store. "Condensed from Aomi's engine architecture doc." POSTER Confirm name use
The awesome list
Six named community projects on the poster — Burns, Ralphinho, Cairo Coder, Agentix, Local Isolated Ralph, VibeAudit — curated at smithersai/awesome-smithers. POSTER WEB
elizaOS
The public cohort profile lists "adoption by elizaOS as a workflow engine" among Smithers' traction. Stated there only; not repeated on the team's own poster. REGISTRY Confirm
What's next
From the poster's what's-next panel.
Forward-deployed
"Smithers in real businesses, one workflow at a time." POSTER
Design partners
The ask, verbatim: "Do you know anybody looking to automate their business with workflows?" POSTER
Open source, always
Free and MIT stays the model; the pitch frames enterprise integration as "the last mile." PITCH Pending approval
In their own words
How Will tells it — from the cohort's mock demo day rehearsal, 2026-07-17, lightly punctuated from a live transcript.
"A completed task is never re-executed." The durability contract, as printed on the team's cohort poster POSTER
"Smithers makes agent workflows reliable. You build them by simply describing them in human language. They run for days or weeks, survive crashes, pause for a human sign-off, and improve the more you use them — and they never lose work."
"The plumbing is more work than the idea. Smithers ships the plumbing as the framework — durability, retries, approvals, observability, all built in. And you don't write the workflows. Agents write them. You just describe the job; your agent builds it."
"The right way to build agents changes every six months. Frameworks built for today's meta will die with it. But Smithers doesn't change. Workflow automation was already big before agents — after agents, it's only going to get bigger."
"The vast majority of users who use Smithers never churn. They report it being one of their biggest unlocks in their agentic journey — up there with releases like Claude 3.5. They report bugs rather than quietly churning."
Try it & links
"Ask any agent to read smithers.sh and install the smithers plugins and skill." POSTER
| Surface | Where | Source |
|---|---|---|
| Website | smithers.sh — "Agent workflows with time travel" | WEB Live · 07-22 |
| Docs | smithers.sh/guide/get-started — bunx smithers-orchestrator init | WEB Live · 07-22 |
| GitHub | github.com/smithersai/smithers — MIT · 336 stars · pushed 07-22 | WEB Live · 07-22 |
| npm | smithers-orchestrator — 198,295 downloads, past year | WEB Live · 07-22 |
| Ecosystem | smithersai/awesome-smithers — the full curated list | POSTER Live · 07-22 |
| X | x.com/fucory — the creator's handle | REGISTRY |
| Demo video | Demo-day video on file with the cohort (2026-07-22) | TEAM |
| Contact | will@smithers.sh | POSTER Confirm preferred |
Provenance ledger
Every load-bearing claim and where it comes from; decisions live in §12.
| Claim | Comes from | Tier | Status |
|---|---|---|---|
| One-liner, problem framing, features, durability contract, time-travel demo | Cohort poster, July 2026 | POSTER | Team's own artifact |
| Will Cory, solo creator; demo-day presenter | Cohort registry; his own pitch opener; top contributor on the public repo | REGISTRY WEB | Corroborated |
| Public numbers (downloads, stars, contributors) | Poster figures re-checked against api.npmjs.org and api.github.com, 2026-07-22 | POSTER WEB | Verified live |
| Open source · MIT | Poster header; repo license field, checked 07-22 | POSTER WEB | Verified live |
| Aomi /build production deployment (Exhibit A) | Poster — names a cohort peer and quotes its engine doc; not independently checked | POSTER | Confirm name use |
| elizaOS adoption | Public cohort team profile only; not on the team's poster or site | REGISTRY | Confirm |
| Prior work (Tevm, EF grant; Guillotine) | Public cohort team profile; star counts there not re-checked, so left off the page | REGISTRY | Self-reported — confirm |
| JJHub lineage (applied as JJHub, week-2 pivot; legacy site) | Cohort records; jjhub.tech fetched 2026-07-22 | REGISTRY WEB | Verified live |
| Links (site, docs, repo, npm, awesome list) | Poster + registry, all live-checked 2026-07-22 | REGISTRY WEB | All live |
| Pitch excerpts (§09) | Mock demo day rehearsal, 2026-07-17; lightly punctuated, two ASR garbles decoded in a hover note | PITCH | Pending approval |
| Demo video | Team's demo-day asset on file, 2026-07-22; no hosted link in the cohort's links sheet | TEAM | On file |
Founder review
Five things before this page goes public.