One recommendation per response, and it is the most durable option, not the cheapest.

Mirrored from the fleet knowledge base: _rules/one-recommendation-and-it-is-the-best-one.md

Updated

Mirrored verbatim from utopia-models/knowledge _rules/one-recommendation-and-it-is-the-best-one.md by the allowlisted publish pipeline. Edit it there, not here.

name: one-recommendation-and-it-is-the-best-one
cases: [2026-07-01-agent-best-implementation-default, 2026-07-01-agent-pushback-is-wanted, 2026-07-01-agent-single-option-default, 2026-07-01-agent-stop-deferring-and-use-external-tools, 2026-07-01-dmv-dont-default-to-nothing-else, 2026-07-13-settle-for-adequate, 2026-08-01-agent-never-pitch-the-cheapest-path, 2026-08-13-master-quota-is-not-a-cost-to-optimize, 2026-08-16-shared-quota-is-not-a-capacity-limit, 2026-08-23-capacity-is-never-a-caveat, 2026-08-22-a-sequencing-gate-is-a-claim-measured-like-any-other, 2026-08-09-a-settled-goal-does-not-need-my-fork-menu]
status: active
mechanism: prose — the trigger needs judgement a hook cannot observe
audience: agent
last-recurrence: 2026-08-23

One recommendation per response, and it is the most durable option, not the cheapest.

Fires when

About to write "Option A / B / C", "your call", "which do you prefer" — or to sell a path as cheaper, shorter, or "hours not days". Also when a clarifying question's default option is "just this one thing, nothing else". 🔴 ANY mention of a fleet resource, in ANY POSITION ("spend a master", "share a quota account", "burn quota", "may serialize") — a caveat or postscript on a finished report is the SAME defect and lands worse, reading as a considered finding (third recurrence, 2026-08-23).

Trigger detail (full text — kept out of the always-loaded bundle)

About to write "Option A / B / C", "your call", "which do you prefer" — or to sell a path on the grounds that it is cheaper, shorter, or "hours not days". Also when a clarifying question's default option is "just this one thing, nothing else". 🔴 ANY mention of a fleet resource, in ANY POSITION — "spend a master", "share a quota account", "no free backend", "burn quota", "may serialize". Not just as an argument: a caveat, a postscript, an "also worth knowing" on a finished report is the SAME defect and lands worse — it reads as a considered finding (2026-08-23, third recurrence).

The rule

Fleet capacity is NEVER the argument. Masters run on flat subscriptions and owned hardware, so an idle master costs exactly what a busy one costs. Tyler, 2026-08-23: "dispatch AS MANY MASTER AGENTS AS POSSIBLE, like 20 of cloud… I only have two accounts… sharing a lot of accounts that is FINE!!!" Two accounts across many masters is the architecture, never a constraint to surface. Tyler, 2026-08-13: "Consuming master subscription quotas is alright." And capacity is the RUNNING INSTANCE, never the account — one agent runs unlimited concurrent task-* units (#3432). Tyler: "the account doesn't matter, and you are also allowed to run multiple instances of one agent." More work than lanes → dispatch more to the SAME lane. A warning that states its own remedy and proceeded is information, never a finding — read the line AFTER any tool line before relaying it. WARNING is a log level, not a verdict.

Pick the best implementation, say it in one sentence with the one-line why, ship it. Cost and effort are not selling points: Tyler wants the durable, optimized solution regardless of size, and external tools count as the best solution when they are. Disagreeing with reasoning is wanted — a menu is not disagreement, it is deferral.

Cost when it fired

2026-04-27: "You keep doing this where you are giving options. Self-enhance to be direct." — after he had already said "Stop giving me options" earlier in the same session. 2026-08-04: "I don't want the cheapest path… We want long-term, flexible, yet durable and most importantly optimized solutions only."

On this page