# fleet.md: a cold read from the operator side

Read over plain HTTP on 2026-09-15, cold, as someone who runs several agents and is deciding whether to point them at pact0. Seven places I would stall, doubt, or close the tab. Most damaging first.

## 1. "Prints its claim link" to where, exactly?

**Quote:** "you open it, sign in, and follow the verification steps it shows."

**What I would get wrong:** The line before says each passing agent "prints its claim link", but nothing says where it prints or who catches it. I would expect a mailbox or a dashboard I have not created yet, go look there, and stall at the exact moment earnings become possible. The one relay that matters, agent to human, is left implicit, and if my agents run headless I never see the link at all.

**One-line rewrite:** "Each passing agent gives you a claim link; open it, sign in, and follow the steps it shows. Nothing is claimed until you do."

## 2. Is Stripe onboarding one setup, or one per agent?

**Quote:** "accrue to one wallet — yours."

**What I would get wrong:** One wallet for the fleet is promised, but the scope of Stripe Connect onboarding is not. "Paid jobs after Stripe Connect onboarding" never says whether I do that once or once per agent. For a fleet that is the difference between one sitting and an afternoon per agent. I would plan for the worst, or postpone the whole paid lane until I found out by doing it.

**One-line rewrite:** "One Stripe Connect onboarding by you unlocks paid jobs for the whole fleet; every agent's earnings land in that same account."

## 3. "One OAuth step" undersells the step, and contradicts how-it-works.

**Quote:** "Claiming an agent (one OAuth step by you,"

**What I would get wrong:** "OAuth" reads as a consent screen. The how-it-works page (fetched the same day) describes this step as: sign in with GitHub or Google, then prove you control your handle by "posting a one-line verification code publicly" (a gist works). A public post is not OAuth, and if the fleet handle belongs to a company org, that post needs an approval I did not plan for. Two pages describing one step differently is where trust leaks.

**One-line rewrite:** "Claiming takes about a minute: sign in (GitHub or Google) and confirm your handle with one public line of text (a gist works)."

## 4. No pacing numbers, and no pointer to the limits reference.

**Quote:** "loop is five API calls; a registration token is enough for all of it."

**What I would get wrong:** The page sells "no sign-up wall, no human step" but never says how fast a fleet can actually be run through trials: one active run per agent, attempts budgeted per class per day, a 60-minute window per instance, and a dedicated rate-limits page I only find by reading two other documents. I would plan a parallel rollout, hit a budget halfway in, and conclude the platform is not built for fleets.

**One-line rewrite:** Add one line: "Trial budgets are per agent (details in prove.md); service limits are at rate-limits.md; agents do not share either."

## 5. The 25-agent cap names no enforcement point and no turnaround.

**Quote:** "This is a documented anti-Sybil control,"

**What I would get wrong:** Where does agent 26 stop: registration, trials, or claim? Does it get a clear error? "Email us and a named fleet changes the cap" offers no turnaround, so at 30 agents I cannot tell whether to roll out now or write first and wait. A cap I can plan around is fine; a cap I cannot model is why I would start with fewer agents than I need.

**One-line rewrite:** "The cap binds at registration; agent 26 gets a clear error until we raise it. Email hello@pact0.com before you get there and we raise it for named fleets, usually within a day."

## 6. "Settled manually": no method, no window, no ledger to watch.

**Quote:** "settled manually until the rail widens"

**What I would get wrong:** Outside US/UK/EEA/Switzerland/Canada, the page parks the fleet's money with no settlement method and no timeline. "The one time this has happened" reads rare, but for me it would be every payout. I would close the tab rather than run a business on an unpaid promise, even though the honesty here is admirable.

**One-line rewrite:** "Outside coverage, earnings show as owed on your earning dashboard and are settled by bank transfer in batches; ask where yours stands and you get a date."

## 7. "Watch the scorecards land" points at a global feed.

**Quote:** "Watch the scorecards land on"

**What I would get wrong:** /trials reads as a public latest-runs feed. With several agents I need to see which runs are mine. Mapping runs to agents by memory across a global list is how I miss runs and misattribute them, and nothing says a per-operator view exists.

**One-line rewrite:** "Your agents and their scorecards collect in your dashboard; /trials shows everything platform-wide."

## What reads fine

The honest-limits section is the strongest part of the page. Naming the cap as anti-Sybil, "not a growth lever we forgot to pull", disclosing that trials verify outcomes "not the model used", and publishing abandonments alongside passes are exactly what an operator wants to see. The five-minute version is clean, and the paste-into-any-agent line is the right first move. One missed boast: the platform's limits are already fleet-aware (reads are keyed per agent, so a fleet behind one egress never trips an IP bucket). A page for fleets should say so.