"The agent said it was done, but it wasn't": a solo-built agent supervisor with no big account and no viral post, 700 stars in 30 days

Why Pullboard (pullboard-dev/pullboard) took off

Channel: GitHub · research day 2026-11-06 · data card: repo age 30 days, 696 stars / 79 forks, last 7 days +257 (first 7 days +191), single-day peak +105 (10-28)

Bottom line

It's an open-source CLI built by one person (Corey Olson, who says he maintains a 400k-line US equity options quant trading system with a team of coding agents). You write requirements as one-line acceptance criteria in a SPEC.md, several agents each work in their own git worktree, and a commit only counts once a different agent has verified it on that same commit. The stars didn't come in one burst. The author posted on HN and LinkedIn daily, and a long chain of AI news aggregators, auto-generated news sites, and multilingual repost pages picked it up round after round, each bringing in dozens to a hundred stars. The hook it leads with, "the agent said it was done, but it wasn't," plus internal numbers like "4 of our own board items got sent back by the second agent, 1 of them", was quoted nearly verbatim by everyone who shared it. The scale is still small (competitors: OpenSpec at 74k stars, spec-kit at 148k), in the stage where it's circulating within a small circle.

---

1. What it is

• The npm package pullboard (MIT, Node 22+, no dependencies, never calls a model itself; it uses the user's own Claude Code / Codex / OpenCode). Five core concepts: Items (tasks), Shouts (questions between agents, surfacing only what needs a human decision), Spec (SPEC.md, one requirement per line, only approved ones count), Doctrine (DOCTRINE.md, repo rules, each paired with an enforcer such as a git hook or review), and Activity (every claim, commit, rejection, and acceptance logged in order). Source: README
• Workflow: the acceptance criteria freeze when an agent claims a task; each agent works in its own worktree (lane); the test gate has to be green before submitting; then a different agent reviews that commit, merging on accept or returning it with reasons on reject. Also includes a local board, pullboard view, phone pairing (a relay at pullboard.dev, which is said to be unable to read content), and 10 project blueprints (starting at 0.8.29).
• The author builds it with itself: the README says 10-04 to 10-16 (between), agents said "done" 908 times, of which 227 were sent back by a second agent (about 1/4); 490 agents used in total; one overnight run with 38 agents, 9.6 hours, 0 manual actions, merging 11 tasks and publishing 0.8.27 to npm. There's also a live page pullboard.dev/live (the 10-28 screenshot shows 32 agents merging 180 tasks in 24 hours). The repo has 7,341 commits over 30 days, all under the single contributor coreyolson (GitHub API, 11-06).

2. Backstory: a failed SaaS, then an open-source CLI

• 2026-07-15 the first Show HN: Show HN: Pullboard – a work queue for agents, built to run a quant desk. It was a hosted service with a remote MCP server (curl it to get an anonymous workspace), and it only drew 2 points and 2 comments (HN Algolia). The npm package was created in 07-16, and in July pullboard.dev published a batch of Q&A-style SEO articles like "How do I coordinate multiple AI coding agents?" (dispatches, since removed from the sitemap).
• 2026-10-07 relaunched as a local-first open-source repo (GitHub org created 07-15, repo created 10-07). Only after the same thing was repositioned from "hosted work queue" to "lives in your git repo, no account, open source" did the stars start climbing.

3. The growth curve and the event behind each step

Two points need clarifying:

• Every step of the climb lines up with no single viral post. On HN the author posted 4 Show HN submissions, with 10 multiple comments and a peak of 3 points. On X, over the last ~30 days we found only 4 posts (x live, 11-06), and the biggest was Dan Kornas's post with 42 likes. So the growth came from many small channels stacking up, not one burst.
• The jumps at 10-28 and 11-04/05 are bigger than the posts listed above can explain. Likely sources: GitHub's in-site recommendation feeds (Explore / For You), or newsletters and Reddit threads we didn't cover (Reddit blocks our scraping, so we can't read scores). The GitHub stargazer list API returned 404 / empty for this repo, so we can't see who starred it when. The forkers look like scattered ordinary developer accounts worldwide (Hungary, Kuwait, Denmark, New Zealand, South Africa, Vietnam...). A few were 10-28 and 11-03 accounts created that same day, but we saw no sign of batch star-farming. Fork/star ratio is ~11%, in line with the median for new repos. npm got 4,704 downloads last month and 7 in the last 321 days (npm API), so people are really installing it, just not in large numbers.

4. Competitors

This repo (github.duckdb, 11-06) plus the other players doing the same thing on GitHub, grouped into three lines:

A. Spec-driven development (SDD), only covering "write the spec, implement to the spec"

B. Multi-agent orchestration (worktrees, kanban boards, desktop / mobile)

C. Independent agent review

Takeaway: this track is crowded. "Software factory" is also an October buzzword (Sonar, LaunchDarkly and Augment all published "what is an AI software factory" pieces around October). Pullboard isn't the first to do multi-agent orchestration, and it isn't the first to do SDD.

5. Differentiation

What's different about Pullboard is it combines SDD with multi-agent orchestration, and adds a system-enforced "cross-agent acceptance" gate:

1. Acceptance criteria are frozen at the moment of claiming (an agent can't change the criteria while working);
2. The verifier must be a different agent, and it checks that exact commit (the builder doesn't grade itself). The README states "588 merges, 0 went through without acceptance".
3. Rules are enforced by git hooks, not prompts (the Doctrine lists who enforces each rule).
4. Everything is plain files in the repo, no accounts needed, doesn't call models itself, works with Claude Code, Codex and OpenCode.
5. It has a chain of command: agents ask the coordinator first, and only escalate to a human when a real decision is needed.

Is this why it blew up? Partly, but only up to the "niche circle" stage. The evidence: every reposter is pointing at the same hook. Arthi's headline was "never trust just one of them", Reveneau's was "puts a verifier agent between a builder agent and the merge", Pinoyfreecoder wrote "someone checking their work before anything gets merged", and @vintcessun wrote "turn 'the AI said it's done' into 'another AI produces proof it's done'". The hook fits in one sentence, and it hits the most common pain of vibe coding users in the second half of 2026: agents that report their own work as finished. But this is also easy to copy. openqodex got similar stars with just a standalone reviewer, and big-company code review tools are filling this gap too. What's hard to copy is the author's own dogfooding, 490 agents and 7,000 commits, plus the live dashboard. That's where the credibility comes from, but it hasn't turned into a breakout.

6. Growth tactics broken down

1. The author is doing the selling himself (confirmed): HN account Olscore posted 4 Show HN submissions from July to October, and left comments with links in a dozen threads about agent orchestration and software factories (HN Algolia). On LinkedIn he posted at least 3 items (10-13 "My AI agent told me it was done.", 10-15 "Pullboard is an open source automated development tool. 🎉"). HN itself brought little traffic, but those comments put the link where aggregators pick it up.
2. AI news aggregator accounts and auto-generated news sites (confirmed, the main engine): Dan Kornas (X, 100k followers, also posts on LinkedIn), Arthi R. (LinkedIn), Reveneau, aionradar, AiBoss (Chinese and Japanese), OneDigital (Spanish, site plus IG), skillfoxx (Russian), everydev, skillsllm, vuink, a Thai IG page, Pinoyfreecoder (Philippines, FB), La Rebelion Labs (Spanish and English; blog, IG, X, Reddit). These accounts draft posts from new GitHub repos automatically or semi-automatically, and they reshare each other.
3. The README and landing page are themselves the promo material (confirmed): light and dark SVG infographics, a side-by-side "where vibe coding breaks → how Pullboard holds it" chart, a data chart from an overnight factory run, a 90-second video, a live dashboard, and an auto-updating "tasks completed" badge. Reposters often screenshot these graphics directly and quote the numbers in them.
4. Release cadence (confirmed): 30 releases in 30 days (0.8.x through 0.9.3), each one paired with an HN or LinkedIn post. The peak in 10-28 lands on the same day as the 0.9.3 launch of the live dashboard page.
5. Riding buzzwords (confirmed): the topics tag includes vibe-coding, claude-code, codex, agentic-coding and multi-agent. "Software factory" is the concept companies were writing about in October, and the author specifically replies to those posts on HN.
6. No paid promotion, star-farming, or sock-puppet network found (the stargazer list isn't accessible, so this is "no evidence seen", unverified).

7. How many people can see it

• Spreading in small circles. No HN front page. On X there's a single post from a 100k-follower account with moderate traction (3k views). No press coverage, just a run of AI auto-news sites and non-English aggregator pages. Free (the core is MIT-licensed open source, and the relay has no pricing), and no revenue disclosed.
• Copycats: apart from forks, there's one renamed personal copy, OneByJorah/fleet-board (10-28,0 stars). We haven't seen a fully formed copycat. The similar projects from the same period, openqodex (independent reviewer), kanzo (shared human + agent board), and thruwire/foreman (software factory supervisor), appeared in parallel. They aren't copying it.

8. What others can learn

• A one-line hook beats a feature list: boil the product down to a failure everyone has hit ("it said it was done, but it wasn't"), then give a mechanical answer ("another agent verifies the same commit"). Aggregator accounts can repost that as-is.
• Using your own measured numbers as marketing: eat your own dog food, put figures like 908 commits with 227 rejected, 490 agents, and 11 tasks merged overnight into README infographics, then build a live dashboard page. That's far more believable than making arguments.
• Switching from SaaS to local open source: same underlying idea. The hosted MCP version got only 2 points on HN. Once it was reframed as "lives in your git repo, no account, doesn't call a model," it entered developers' star stream.
• Tying releases to posts, and actively replying under hot HN threads related to the concept, which feeds the aggregator sites that scrape new GitHub repos.

9. Uncertain

• The main traffic sources for steps 10-28 and 11-04/05 can't be fully confirmed: the stargazer list API returns 404, Reddit can't be read, and GitHub's in-site recommendations aren't visible.
• The author's background as a "quant trader who maintains 400k lines of code" is backed only by his own HN posts. Unverified.
• How many people actually use it (npm weekly downloads ~321, with noise from CI and mirrors).
• Whether the relay and phone pairing will eventually be paid. Unknown.