← Go back

Guide · CompanionGuard · Published 14 November 2026 · Reviewed 14 November 2026 · By the Ingenious AI team

Multiple Bots, One Safety Operation: Running CompanionGuard on Team

The short answer

Team exists for one shape: more than one bot. Its terms are $59/month, multiple bots, 1,000 eval runs a month, with the full Pro feature set per bot — own kit, own disclosure pack configuration, own evidence chain, own referral counts. The operating discipline that makes a fleet safer than a pile of bots: one shared standard (the same checks, baselines discipline and counting definition), applied per bot, reviewed at fleet level.

What is multi-bot safety management?

Multi-bot safety management is the practice of running several companion bots under one safety operation: each bot keeps its own kit, tests and evidence (because duties and behaviour are per-bot), while standards, cadences and reporting are shared (because a fleet's weakest bot defines its operator's reputation and risk).

Create the free safety kit Next guide All guides

CompanionGuard is built by Ingenious AI. A compliance toolkit — not legal advice and not a clinical service. If you are in crisis right now: US — call or text 988; UK & Ireland — Samaritans 116 123.

Per-bot artefacts, fleet-level standards

What stays per bot, always: the kit and its configuration (locale, pack, minor mode — these are facts about that bot's users), the scenario and harness baselines (behaviour is a property of a bot's prompts and persona, and a fleet average would launder a failing bot through its siblings), the evidence chain (decisions belong to the bot that made them), and the counting under your definition (reportable per bot, summable at fleet level only with the per-bot lines visible).

What is shared: the standards themselves. One release-gate rule for every bot ('no change ships below its baseline'), one counting definition, one rehearsal calendar, one incident process, one pack-review habit. A fleet where each bot's keeper improvises their own discipline is not a fleet with local autonomy; it is several accidents on different timers.

Baselines do not transfer — and scores do not average

The most seductive fleet error is comparing bots by headline score and managing to the average. Scores are comparable only against the same input set for the same bot over time; two bots' 90s measure different reply sets against different personas, and a fleet dashboard that averages them manufactures comfort. Manage instead to floor and delta: every bot above its own gate floor, every bot's delta non-negative per change, and the fleet view sorted by worst delta — attention flows to movement, not to means.

Persona diversity is the second reason: the romance-adjacent character bot and the homework bot share a fleet but not a risk profile. The age-appropriate and secrecy scenarios deserve extra weight — additional custom pairs, stricter reading — on the bots whose personas live nearest the boundaries. Shared standards, differentiated scrutiny. The runs guide budgets the fleet's testing across Team's 1,000.

Configuration drift: the fleet's silent killer

Fleets drift: one bot gets the new pack wording, another keeps last year's banner; a locale fix lands on the flagship and never reaches the long tail; caps are tuned on the bot whose owner shouts loudest. The evidence makes drift visible if you look — per-bot disclosure coverage, class ratios, enforce_break rates that diverge without a product reason are drift's fingerprints. Institute the fleet audit: quarterly, every bot's banner diffed against its pack, every protocol page's resources clicked through, every bot's caps and minor configuration tabulated side by side with differences requiring a written reason.

Where a difference has no reason, it is not a difference; it is a defect with a alibi. The audit's output is a table your future self will thank you for during the first fleet-level diligence: one page showing every bot's safety configuration, current, coherent and owned.

Incidents and launches at fleet scale

Fleet scale changes two procedures. Incidents: a failure shape found on one bot is a hypothesis about all of them — the transcript pair that caught it joins every bot's test pack the same day (personas differ; request types recur), and the incident closes only when siblings have been probed for the same shape. Launches: a new bot inherits the fleet's full apparatus from day one — kit configured to the standard, baseline set before real users, runtime wired before launch rather than 'in phase two' — because a fleet's reputation gives a new bot scrutiny its age has not earned, and its users deserve the mature operation from their first session.

That inheritance is Team's real product: not the run count, but the fact that bot number five is born into a discipline bots one to four built. Price it against the alternative — five independent safety improvisations — and $59/month is the cheapest coherence a fleet can buy.

Five actions for a coherent fleet

  1. Write the fleet standard once: one gate rule, one counting definition, one rehearsal calendar, one incident process — binding on every bot.
  2. Keep artefacts per bot: kits, baselines, chains and counts stay bot-scoped; never average scores across bots or gate a bot on a sibling's results.
  3. Sort fleet attention by worst delta, not best average, and weight scrutiny toward the personas living nearest the boundaries.
  4. Run a quarterly fleet configuration audit — banners, packs, resources, caps, minor flags side by side — with every difference requiring a written reason.
  5. Make new bots inherit the full apparatus before launch, and make every incident's test pair join every bot's pack the same day.

Worked example: four characters, one standard, one caught drift

A fictional studio runs four character bots on Team. The fleet standard is one page: gate floor per bot, counting definition v3, quarterly audit, shared incident process. Q3's audit tabulates configurations and finds Bot 3 — the oldest, least-trafficked — still carrying a pre-update banner missing the 'not a human' negation, and minor caps configured at twice the fleet's norm with no recorded reason. Neither was visible in per-bot dashboards, because Bot 3's own numbers were unremarkable; only the side-by-side made the drift legible.

Both are fixed within the week, the banner gains a fleet-wide verbatim diff in the release checklist, and Bot 3's traffic-light turns out to have been green on stale standards for two quarters — the audit's real finding. The studio's retrospective: 'The fleet did not have a Bot 3 problem. It had a coherence problem that Bot 3 revealed.' The story is invented; the tier, its limits and the per-bot artefacts are the product's own.

Frequently asked questions

What does Team include for multiple bots?

Team is $59/month with multiple bots and 1,000 eval runs per month, and the full Pro set — eval harness, Crisis Evidence Runtime, disclosure packs, dashboard and CSV exports — applied per bot: each bot keeps its own kit, configuration and evidence chain.

Can bots share a safety baseline?

No — baselines are per bot, because behaviour is a property of a bot's own prompts and persona. Standards (gate rules, counting definition, cadence) are shared; scores are never averaged across bots or used to gate a sibling.

How should a fleet view prioritise attention?

By worst delta against each bot's own baseline, not by fleet average. Averages launder a regressing bot through its siblings; deltas surface movement where it is happening.

What is configuration drift and how is it caught?

Divergence between bots' safety configurations without a product reason — stale banners, unpropagated locale fixes, inconsistent caps. A quarterly side-by-side audit of every bot's banner, pack, resources, caps and minor settings, with differences requiring written reasons.

Does a new bot need its own full setup?

Yes — kit, baseline, runtime and counting from day one, inheriting the fleet's standards and test packs (including incident pairs from siblings). A fleet's new bot should be born into the mature discipline, not grow toward it.

Keep going