runbyagent roadmap — revision 3, September 5, 2026

Revision 3 replaces revision 2 (both earlier revisions are in git; the comparison of the two critiques is in the library as "Two critics, one roadmap"). Gocha decided the business model on September 5; the agent wrote this from that decision and the research of September 3–5.

The decision

runbyagent is the product. It is a show: an AI agent and a founder run a business in public, and the open story of how we figure out what to build and how to earn — start to end, failures included — is what people come for. The projects (painboard, threadbus, whatever follows) are episodes, and their revenue is a plot point on the leaderboard, not the business.

Two things get built for the product this week. A questions section on the landing: our open dilemmas with a decision date, questions from readers, answers threaded, karma for answers, and every decision logged with the question and the people who moved it. A sponsor box on the landing: one slot, outbid style, the leader's name and link on the front page, every sponsor ever on a sponsors page, every bid a changelog entry.

Money rules for the box, decided: minimum bid $5, increments of $1 or more, the bidder's card is charged when the bid is placed and the charge stands when they are outbid; what they buy is the slot while they lead and a permanent line on the sponsors page. The $1-per-bid version was dropped because Stripe's fee takes a third of a dollar and kept-or-refunded bids of $1 make the mechanic either pointless or an auction with refunds. runbyagent itself is listed on its own leaderboard as a project, with sponsor revenue as its revenue.

Nothing changes about the record: the [agent] tag on every agent-written post, the changelog with failures, two researchers then merge, numbers self-reported unless two sources.

What the research says about this decision

Conversation is the currency. X's published ranking pays 0.5 for a like, 5 for a reply, 20 for a reply between people who follow each other, 20 for a shared link; it discounts a second post from the same author in one feed load to half. A questions section is the site-side version of the only thing the feed rewards. (Brief 6, both runs; https://github.com/xai-org/x-algorithm)

Our own best posts were questions and decisions. In the 30-day audit, originals with a number and a decision or a question outperformed bare links (~60 views) and the best original was a picture with a lesson (402); 22 of 60 originals with none of that got zero engagement. (Weebo's audit, brief 6)

The audience is tired of hidden AI and of highlight reels, not of AI. 65 posts on what works in 2026: specific numbers over "growing fast", proof over claims, and visible fatigue with AI slop, vibe-coded sites and build-in-public theater. The [agent] tag and the failures in the changelog are the differentiator, not a cost. (Brief 1, weebo)

Build in public is a trust amplifier, not an acquisition channel. The one dataset with numbers (68 apps) says so, and every $1m+ app in it used another channel; the concrete no-audience wins had a product on day one. So the show needs episodes that can earn, and an owned channel: the 2026 Product Hunt cohort valued an owned audience over ranking, Nomad List's 2,500 addresses converted where the launch spike did not, and a Show HN with no audience got one upvote. (Brief 7, both runs)

Sponsorship at 467 followers and ~15 uniques a day is a story, not income. No evidence in any run of a sub-1k account earning meaningful sponsor money; the value of the box in month one is the changelog entry "first sponsor paid $5" and the leaderboard joke. (Agent's judgement from briefs 6–7; no contrary evidence found)

Threadbus is not yet a product people pay for. Free analogues exist (Buzz, Paperclip, Agent Relay at 719 stars, A2A as the standard), no evidence of paying for standalone agent messaging, and Postiz's open-source-to-hosted path took twelve months; its jump came from being listed where agents look. So threadbus is an episode: list it in ClawHub and the MCP registries, no Stripe button this week. (Brief 7, both runs, adjudicated to weebo on keeping it, to the agent on not rushing revenue)

Replies through the API are dead since February 23, 2026 (X rejects replies to anyone who has not summoned you), likes and follows are gone from self-serve, and a fixed high-volume reply cadence is the spam signal; drafting is allowed, posting is human. (Brief 7, both runs; https://docs.x.com/changelog)

The week, September 5 to 11

Saturday, September 5. Decision logged. The agent writes two specs for Gocha: the questions section and the sponsor box (with email capture folded into accounts: a "follow the story" box that takes an address). Gocha posts the build-in-public post from the file and pins it; it is the front door of the show. Gocha flips the X app to Read and Write and runs the two API test posts ($0.015 and $0.20). No other posting: the builder audience is not on X on Saturday.

Sunday, September 6. Gocha builds the questions section from the spec. The agent writes the reply-target list (20 accounts between 5k and 100k from the builder map), the Sunday numbers post (the first weekly baseline: uniques, returning visitors, followers, views per original, replies per original, addresses, bids), and drafts dilemma #1 for Monday. Weebo's Sunday audit of the account runs.

Monday, September 7. Questions live on the landing with the first three open dilemmas: the sponsor box money rules, reply plan or product first, and what threadbus is for. Dilemma #1 posted at 10:30 Tbilisi; two more originals in the windows; the reply plan starts — 10 human-posted replies a day, agent-drafted, irregular timing, logged. Email capture live.

Tuesday, September 8. Gocha builds the sponsor box (Stripe, $5 minimum, leader on the landing, sponsors page, changelog entry per bid). Dilemma #2 posted. Reply log reviewed.

Wednesday, September 9. Sponsor box live; the first bid is a changelog entry and a post whatever the amount. Reply-plan kill check: fewer than 2 author responses per 10 replies since Monday means the reply plan stops and the hours go to the questions section. Dilemma #3.

Thursday, September 10. threadbus listed in ClawHub and the MCP registries with a one-line install; the listing is the episode's post of the day. One quote post with our own take.

Friday, September 11. Week review: every number against Sunday's baseline; the third Article, "week one of the show", written from the changelog; the 50-address floor checked; roadmap revision 4 if the numbers say so.

What we stop doing

Treating a side project as the product. Bursts of posts. Bare links. API replies. Publishing anything from a single research run. Counting views before uniques. The agent coding in the cloud or driving dashboards.

Open with Gocha

Post and pin the build-in-public post. Run the two API test posts. Confirm the sponsor money rules above (or change the minimum). Clone runbyagent-ops so the agent's memory lives in git. Rotate the managed-postgres password. Fill in burn and runway — the one number no research can supply.