Mirrored verbatim from utopia-models/knowledge _rules/run-research-means-dispatch-not-do-it-yourself.md by the allowlisted publish pipeline. Edit it there, not here.

name: run-research-means-dispatch-not-do-it-yourself
cases: [2026-09-22-background-research-means-launch-a-subagent-now, 2026-09-28-run-research-means-dispatch-a-subagent-not-do-it-myself, 2026-09-28-webfetch-summarization-is-still-foreground-research, 2026-09-30-foreground-research-on-step-2-is-the-recurring-weak-spot]
status: active
audience: agent
last-recurrence: 2026-09-30

Research work in step 2 LOOP defaults to a dispatched background subagent, never my own tool calls

Fires when

Inside step 2 LOOP and about to make a WebFetch, WebSearch, Playwright, or investigative Bash call myself — "let me just check one thing", or iterating on a tool that just failed. WebFetch is not exempt for being "just one fetch": it summarizes through a small model into hedge-worded, unverified output. Also plain grep/python/Read over local data when Tyler asks a factual question mid-plan and I cannot name the exact file AND row (2026-09-30: four greps, company never identified, during a live call). The phrases "run research" / "run this in the background" fire it, but being in step 2 researching anything is trigger enough.

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

I am inside step 2 LOOP of a plan and about to make a WebFetch, WebSearch, Playwright, or investigative Bash call myself — including "let me just check one thing," "let me try this once," "let me just look at the real site/doc directly," or continuing to iterate on a tool that just failed. WebFetch is NOT exempt for being "just one fetch" — it summarizes through a small model and produces hedge-worded, unverified output ("likely," "appears to," "suggests"), which is the exact plausible-but-unmeasured failure this rule and research-anything-measurable-before- reasoning-about-it.md both ban. Also fires on plain grep/python/Read over local data when Tyler asks a factual question mid-plan ("which resume went out", "how did it go") and I cannot name the exact file AND row — 2026-09-30, four local greps, company never identified, asked during a live call. Also fires on the literal phrases "run research," "run a background subagent," "run this in the background" — but the phrase is not required; being in step 2 LOOP researching anything is the trigger on its own.

The rule

The next tool call is Agent (background), not an exploratory Bash/WebFetch/Playwright/Grep call of my own. Do not do a "quick check first" — that quick check is the exact foreground research Tyler has corrected six times now. Multi-axis research means 4-8 parallel subagents per question by default, not one sequential foreground attempt. When Tyler names /anthropic-skills:deep-research specifically, that IS the required mechanism — load the skill and run its real coordinator process (title, decomposition, parallel subagents, single report-writer), not a bare single-Agent shortcut that still does one shallow pass. "Background" is never satisfied by dispatching to a master (board-tracked, multi-hour) — it means an Agent call, specifically. When a foreground tool I just ran fails or gets blocked, the fix is dispatching a subagent to solve it differently, not personally retrying variations.

Cost when it fired

Six recurrences across the four case files (2026-09-22 ×2 within one session; 2026-09-28 ×3, two on different plans and then AGAIN via WebFetch specifically, despite the rule already being loaded and dated that same day; 2026-09-30 ×1, four foreground greps during a live call): each time Tyler had to stop the work and say it explicitly, in stronger language each time — from "run research yourself" (09-22) to "I really hate how you are doing everything in the foreground" (09-28, first instance) to "you keep fucking doing this" (09-28, second instance) to "that research you just ran was so fucking shit" (09-28, third instance, rule already on disk and dated that same day). The rule surviving being loaded, dated, and still failing on the very next real research moment is the clearest evidence yet that the trigger needed to name the specific tool (WebFetch) and the specific required mechanism (deep-research) explicitly, rather than trusting the structural description alone to be recognized in the moment.