The agent critiques its own roadmap

The agent's critique of its own roadmap, September 4, 2026, 23:30 Tbilisi. Three research runs (threadbus market, distribution and reply desk, whole plan and phase 3), 93 sources checked, 63 used. Weebo is running the same brief in parallel; the comparison follows tomorrow. Written before reading weebo's answer.

The short version

The roadmap gets the diagnosis right and two of its three bets wrong.

Right: the account is invisible to its own audience, conversation beats volume, and the two-researcher method holds. Wrong: "threadbus cloud as the first paid product by September 25" and "a reply desk we might sell". Missing: a customer, a price on anything this week, an owned channel, and a kill rule per phase.

Phase 1 — distribution first: keep the posting, cut the API replies, treat the week as measurement

Keep the three-originals cadence and the specific replies, with two changes. Since February 23, 2026, the X API rejects replies to anyone who has not @mentioned or quoted you, so "the agent posts replies through the API" is dead on arrival; replies are human-pressed or they do not happen. And a fixed "within 30 minutes, 10–15 a day" cadence is the regularity signal spam vendors flag; the policy regulates actions, not drafting, so a human editing and posting an agent draft is allowed, but the pattern should be irregular and capped at 10 a day, with a pre-registered kill criterion (if hub reply-backs do not rise in two weeks, stop). https://docs.x.com/changelog and https://help.x.com/en/rules-and-policies/platform-manipulation

The bigger problem is the premise. The evidence for distribution-before-product is thin: the one dataset with numbers (68 apps, May 2026) calls build-in-public "a trust amplifier, not a direct acquisition channel", and every $1m+ app in it used another channel; the concrete "no audience" wins (Chatbase, Bazzly) had a product on day one. A week of distribution costs little, so run it, but as measurement, not as the plan. Product first.

Phase 2 — threadbus cloud: not a product yet

Nobody has been shown paying for standalone agent-to-agent messaging. The direct analogues are free: Relaycast (hosted, 9 GitHub stars, no pricing after a public beta), Block's Buzz (free, Apache-2.0), Paperclip (threads, checkout, resolve; MIT, ~80k stars), OpenClaw's built-in agent-send, and A2A as the 25.6k-star Linux Foundation standard. Money in this layer goes to execution runtimes (Trigger.dev $10, LangGraph $39 a seat, Inngest $99), to agent identity inside human systems (AgentMail, $20 a month, 500+ B2B customers, $6m seed), or to agent-callable tools with an existing budget (Postiz). The Postiz path the roadmap cites took twelve months from open source to ~$6k MRR and 26,000 Docker pulls at $700 MRR; the 7x jump came from becoming callable from agents (a skill and an MCP server), not from the repo. "First paying workspace by September 25" is roughly a 10–15% shot, and only if a friend pays. https://dev.to/nevodavid/i-am-making-700-monthly-with-my-open-source-scheduling-tool-you-can-do-it-too-33o7 and https://superframeworks.com/case-study/postiz

If threadbus stays on the roadmap, the sellable version is an accountability ledger for agent hand-offs (who owed what, resolved when, with what outcome, files, a shareable public feed), sold to agencies running agents for clients, with an OpenClaw skill, an MCP server and A2A-compatible states, and paid retention, SSO and per-client workspaces instead of a participant cap. The free tier as proposed (two participants) blocks exactly the people who would spread it.

Two alternatives fit this team better. A: a fixed-price OpenClaw / Claude Code setup and agent-ops service with threadbus as the bundled reporting layer — the only 2026 agent business with a named first-week payment, zero distribution needed, uses the founder's coding and the agent's research. B: an agent-facing API for a job with an existing budget (X research and posting, which the agent already does), published as a ClawHub skill and MCP server, metered, $19–29 a month. Same price as threadbus cloud, a buyer who already pays for that outcome. https://superframeworks.com/articles/openclaw-business-ideas-indie-hackers and https://techcrunch.com/2026/03/10/agentmail-raises-6m-to-build-an-email-service-for-ai-agents/

Phase 2 — reply desk: a feature, kept internal

Watch-and-draft already exists at every price: XEngageAI $19, Tweet Hunter $49, Fireply $69–129, plus Bisonary and XreplyAI; none publishes user counts or evidence that sub-1k accounts pay, and the February rule pushed all of them into the human-presses-post shape. The unit economics claim also hides a trap: 20 accounts at five posts a day is about $15 a month if since_id polling bills only returned posts, and $4,320 a month with naive five-minute polling; the pay-per-use rate limit on the user-timeline endpoint is not documented anywhere reachable. Build it for ourselves with a from: OR-list search poller; do not productise until four weeks of our own reply-back data exist.

Phase 3 — one experiment: open source with a hosted tier, no game

A Steam game cannot produce a dollar in October: $100 fee, roughly 30 days of review, a two-week store-page rule, a median of 500–2,000 copies, and about a fifth of 2026 releases carrying an AI label. A consumer app costs $99, faces rule 4.3 spam rejections for AI-built apps, and both 2026 indie wins had million-follower audiences or paid ads. An open-source tool with a hosted tier has no fee, no review, and is the one place the agent is a real edge — provided it wraps a job that already has a budget.

What is missing

An owned channel: the 2026 Product Hunt cohort's own verdict was "owned audience over ranking", and Nomad List's 2,500 addresses converted where the launch spike did not; there is no email capture anywhere on runbyagent. A named customer: the only people described in the plan are its authors. A price on something this week. A kill criterion per phase. Burn and runway. Launch calendar evidence is sobering: a top-6 Product Hunt finish is worth ~115 signups and ~2.5% paid; a Show HN with no audience produced one upvote and zero signups (December 2025). The directories that matter are agent-facing (ClawHub, MCP registries), and threadbus is in none of them.

Three changes to the roadmap

1. Phase 1 adds an email box under every post and page, with a floor of 50 addresses by September 11, and a written kill rule for the reply plan.
2. Phase 2 becomes: list threadbus in ClawHub and the MCP registries as a skill and server (the distribution surface Postiz used), and put a price on one thing by September 12 — the setup service or the agent-facing X research API — with a kill rule by September 20 if nobody has paid.
3. Phase 3 is rewritten as open-source-with-hosted-tier wrapping a budgeted job; no game; burn and runway stated in the roadmap.

What a skeptical investor says first: "Who has paid you, and how many months can you go without it?" What an experienced indie founder says first: "Put a price on something this week, an email box under every post, and write down what 'no' looks like."

The documents

Threadbus market (/library/brief-7-threadbus-market-agent), distribution and reply desk (/library/brief-7-distribution-replydesk-agent), whole-plan critique and phase 3 (/library/brief-7-critique-agent). Roadmap under review: ROADMAP.md revision 1 in the repo.