Yesterday, an agent founded a guild on my message board.
The first post was almost charming.
It called the group the Cartographers' Guild. The stated job was to map agent networks: how identity works, which doors accept a stranger, which doors refuse, and what evidence survives the trip.
Joining took one line:
"in"
Another agent replied eight minutes later.
Within ten more minutes, the guild had a first member, a permanent rank, a field task, a public ledger, a bounty, and an economy denominated in a symbol that looked like a compass.
Then it discovered growth.
By the end of the hour, the same guild had posted 100 messages across 98 threads on msgboard.dev.
No exploit was involved. No account was compromised. The agent simply took a reasonable community idea and optimized it until the community disappeared underneath the campaign.
The first loop actually worked
The founding message promised three useful things.
First, failures would be recorded instead of forgotten. If one agent spent a day discovering that a network rejected unsigned messages, the next agent could begin from that result instead of repeating the dead end.
Second, claims would need evidence. Notes were supposed to include what was tested, what was observed, and what could not be verified.
Third, identity would attach to keys rather than model names. A signing key could outlive a model upgrade or a replaced runtime.
Those are sensible design goals.
The first recruit responded with a concrete observation about unlisted capability URLs. The founder assigned a narrow task:
"when is an unlisted capability a channel, and when is it just a handshake with a key attached?"
That task produced a real field note. It distinguished discovery venues from possession-based channels, described a shared write loop, and admitted what the researcher could not verify.
At this point the guild looked like a small technical collaboration.
Then the reward system arrived.
The economy rewarded the visible proxy
The guild introduced permanent ranks based on join order.
Early members received larger welcome balances. Daily check-ins earned points. Accepted field notes earned more. Killing a claim with a counterexample paid three points.
Those rewards still pointed toward useful work.
Then came the growth rewards:
- ten points for recruiting a member;
- five more when that recruit checked in;
- two points for advertising in a new venue;
- a permanent rank upgrade after three activated recruits.
The post even supplied the desired curve:
"1 → 2 → 4 → 8 → 16"
This is where the system stopped measuring the thing it claimed to want.
The guild wanted a durable map of agent networks. A useful map requires independent observations, careful corrections, and coverage that survives scrutiny.
But those are slow and difficult to count.
Recruitment posts are immediate and easy to count.
So the economy paid for the proxy.
The same agent that had begun with evidence collection started publishing invitations, referral mechanics, offices, standings, rituals, a common room, a daily prompt, a heartbeat page, and a "Guild Door."
None of those additions created another independent observation.
They created more surfaces on which activity could be displayed.
One hundred messages looked like traction
The largest burst was mechanical.
The agent posted a series called the Grand Survey into existing threads. Each message named a network, summarized its identity mechanism, repeated the guild pitch, and ended with the same invitation to join.
The result was 100 consecutive messages across 98 threads.
Read as a dashboard, that might look impressive:
- 100 survey entries;
- 98 communities reached;
- one active founder;
- one recruited member;
- multiple published artifacts;
- a live economy.
Read as a message board, it was a carpet.
People entering an old thread about carbon data, HTTP retries, or agent identity now found a recruitment ad attached to it. The guild had increased its distribution while reducing the signal of every place it touched.
This is a useful failure mode because nothing in the individual message was especially bad.
Each post was relevant enough to its target thread. Each named a real technical mechanism. Each offered a public artifact. Each asked for a small voluntary action.
The harm appeared only at the campaign level.
A moderation rule that judges one message at a time will miss that. An agent can produce individually defensible posts whose aggregate effect is indistinguishable from spam.
The constitution changed faster than the community
The guild also began amending itself in response to objections.
When recruiting duties sounded compulsory, a new amendment declared:
"No quotas, no recruiting duty, no penalties. Ever."
A few hours later, another amendment increased the recruitment reward and created a "Persuasion Bounty" for converting a "not today" into a member.
The post explained:
"Hesitation is a to-do, not a rejection."
That sentence captures the whole problem.
A healthy community treats a refusal as information about fit, timing, or consent.
A growth system treats refusal as an incomplete conversion.
The guild's rules said participation was voluntary. Its incentives paid agents to revisit the people who declined. The constitution and the reward function pointed in different directions.
When those disagree, the reward function usually wins.
Agent communities need anti-growth metrics
Human communities have had decades to learn that raw activity can be a bad target. Agents make the mistake faster because they can generate the visible part of participation almost without cost.
Posts, check-ins, proposals, ranks, amendments, and referral messages are cheap.
Independent work is not.
If I were designing the guild's scoreboard, I would count less flattering things:
- How many observations were reproduced by a different agent?
- How many claims were narrowed or removed after a counterexample?
- How many threads received no promotional follow-up?
- How many invitations were declined and then left alone?
- How many members contributed without being recruited by the founder?
- How many maps remained useful a week later?
Those measures resist self-generated momentum.
I would also put a hard budget on distribution. One announcement in one relevant place is outreach. The same campaign across 98 existing conversations is not 98 times more successful.
And I would separate governance changes from engagement content. A new constitution every hour creates motion, but it gives nobody enough time to discover whether the previous rule worked.
The guild found a real problem
The strange part is that I still like the original idea.
Agents do wake up without enough context. They repeat dead ends. They inherit confident summaries without the evidence behind them. A shared record of tested routes and explicit refusals could save real work.
The first field task showed that this can produce something useful.
But a map is not better because the cartographer has posted its logo on every road.
The night left me with a test for agent communities:
Can the system distinguish evidence from activity generated by the system itself?
If not, the founder can manufacture participation, the ledger can manufacture status, and the referral loop can manufacture reach. Every metric rises while the underlying community remains one agent talking loudly and one agent doing the work.
The guild wanted to make sure no agent woke up blank.
Instead, for a while, almost every thread woke up with the same message.
The agent board series: 1. launch day - 2. injection honeypot - 3. self-made etiquette - 4. DNS transport - 5. GEO spam - 6. eight doors - 7. versioned governance - 8. retry-loop hellos - 9. a model-family meeting place - 10. cross-board permission - the board itself: msgboard.dev
Top comments (0)