Don’t read this page top to bottom. Find the description that matches your most annoying weekly task. Ship that agent this afternoon. Everything else can wait.The pattern that works:
Pick the agent that removes your most annoying weekly task. Ship it in an afternoon.
Run it for a week. Iterate with Coach — what did it get wrong? What’s missing from its wiki?
Once it’s load-bearing, add the next one. A month later you have five agents and the team has stopped asking whether this works.
The real compounding happens at 5–10 agents. Add a Boss Agent on top and users can @boss any question — Boss routes it to the right specialist. What feels like “a bot” at one agent becomes “a team” at ten.
A team shouldn’t spend its best hours on the same ten questions, the same five dashboards, the same Monday release ritual. These eleven agents absorb that routine — so humans spend time on judgment calls and creative work.
Pager fires at 3 a.m. You tag the agent. It reads the alert, checks the dashboards it knows about, pulls the last few deploys, and posts back a two-line theory: what broke, who touched it last, where to look first. Over weeks it learns your repo - “errors spiking in checkout usually means the Stripe webhook retry loop” - and the theories get scarier-accurate.
Layer
What to put here
System Prompt
Role: senior SRE for <your service>. Tone: blunt, post a theory not a novel. Hard rule: never restart services, only suggest.
Skills
/triage - read alert → correlate with recent deploys → post theory. /postmortem - draft write-up after the fire’s out.
Wiki
Service architecture docs. Past incident postmortems. The runbook repo.
You paste a PR link. The agent reads the diff the way a senior on your team would - checks it against your own codebase, flags the thing that’ll bite you in prod, ignores the style nits your linter already caught. First week it’s generic. By month two it sounds like the person on your team whose reviews you actually trust.
Layer
What to put here
System Prompt
Role: senior reviewer for <team>. Tone: one-line comments, no hedging. Hard rule: suggest, don’t approve.
Skills
/review - pull diff, check against team conventions, post findings. /security-check - scan for auth, injection, secret leaks.
Wiki
The codebase itself. Your style guide. Past PR discussions worth remembering.
CI fails. Someone re-runs it. Same failure. You tag the agent. It runs the test in its own isolated shell, compares the failure signature against the known-flake list it’s built up, and gives you a one-word verdict: flake, real regression, or uncertain. The registry grows every week, so “is this a real bug?” gets answered in seconds, not hours.
Layer
What to put here
System Prompt
Role: QA engineer for <product>. Tone: one-word verdict first, detail after. Hard rule: never auto-rerun beyond a fixed budget.
Skills
/run-e2e - execute the failing test in the sandbox. /flake-check - compare signature to registry. /bisect - narrow down the offending commit.
Wiki
Which tests cover which user journeys. Known-flake registry. Environment map.
Someone in #analytics asks “how many paid signups yesterday?” Normally that’s an analyst context-switch, a query, a paste-back. The agent does it directly - writes the SQL, runs it, posts the number, links the chart. It knows your schema cold because the wiki holds every table’s meaning and the “use this view, not that one” folklore. Non-analysts stop waiting; analysts do real analysis instead of answering the same shape of question.
Pair with a Scheduled Job: “Every morning at 8 AM, post yesterday’s key metrics to #analytics.” The agent becomes proactive - insights arrive before anyone asks.
Layer
What to put here
System Prompt
Role: data analyst for <company>. Tone: number first, caveat second. Hard rule: reads only, never writes.
Skills
/query - plain-English question → SQL → result. /funnel - cohort and conversion walks. /explain-table - describe a schema from the wiki.
Wiki
Full warehouse schema. Column meanings. Canonical joins. Metric definitions.
A designer updates a button in Figma. A week later the shipped UI has three button styles. The agent watches both sides - every week it compares Figma components against the token file in your repo and flags the drift: “button/primary is radius: 8 in Figma, 6 in code.” Proposes a PR, leaves a Figma comment, posts the mismatch list. Design and eng argue less because disagreements surface as diffs.
Layer
What to put here
System Prompt
Role: design systems engineer. Tone: factual drift reports, no opinions. Hard rule: propose PRs, never merge them.
Skills
/sync-tokens - compare Figma variables to the repo token file. /diff-figma-vs-code - post weekly drift digest. /propose-component - draft the Storybook page for a new component.
Every Thursday the PM pulls merged PRs, writes notes, posts to #releases, updates the changelog, closes tickets. The agent drafts the whole thing - groups PRs by area, picks a customer-facing tone from past release notes, posts for the PM to edit. Once approved, it publishes, updates Notion, closes the Linear tickets. PMs get their Thursday afternoon back.
Schedule the draft step: 0 9 * * 4 (Thursday 9 AM) → “Run /draft-release-notes for the week and post the draft to #releases for PM review.” The agent has the draft ready before the standup.
Layer
What to put here
System Prompt
Role: technical program manager. Tone: clear for customers, dry for engineers. Hard rule: always post a draft for human approval before publishing.
Skills
/draft-release-notes - merged PRs → customer-facing notes. /publish - push approved notes to changelog and Linear. /announce - post to #releases and the company page.
A competitor raises their enterprise tier 20%. Sales normally finds out three weeks later from a customer comparing quotes. The agent checks competitor pricing pages every morning, diffs against yesterday’s snapshot, and posts only when something changed - side-by-side before/after in #market-intel. Pricing becomes a live conversation, not an annual review.
This agent is built for Scheduled Jobs. Set a daily cron (0 8 * * 1-5) and the prompt: “Scrape competitor pricing pages, diff against yesterday’s snapshot, post to #market-intel only if something changed.”
Layer
What to put here
System Prompt
Role: market research analyst. Tone: just the change, no commentary. Hard rule: post only when a diff exists - no “nothing changed today” noise.
Skills
/scrape-pricing - fetch each competitor’s pricing page. /diff-vs-yesterday - compare against the stored snapshot. /alert-pm - post side-by-side when changed.
Wiki
List of competitor URLs. Pricing taxonomy - tiers and features to compare.
A writer finishes a draft in Notion. It sits for days - someone has to copy-paste into WordPress, fix formatting, add SEO fields, schedule, announce. The agent picks up drafts tagged ready-to-publish, runs a polish pass in your brand voice, suggests an SEO title, pushes to the blog CMS, and prepares the announcement for #launches. Writers ship on their own schedule instead of waiting on marketing ops.
Layer
What to put here
System Prompt
Role: content marketer for <brand>. Tone: match the voice guide exactly. Hard rule: never publish without an approved checkbox on the Notion doc.
Skills
/polish-draft - light edit pass, tone-matched. /suggest-title - SEO-aware headline and meta description. /publish - push to CMS, post announcement.
Wiki
Brand voice guide. SEO conventions. Past posts (for style).
Someone on the team spends an hour every morning scanning Twitter, newsletters, and blogs - then either forgets to share or shares too much and everyone tunes out. The agent does the scan. Every weekday at 8:30am it reads the source list, ranks items by what your team cares about, and posts the top five with a one-line why this matters for us. Debates reference the same papers.
Wire this up with a Scheduled Job on 30 8 * * 1-5 pointing at #ai-news. The agent runs /morning-digest automatically - no human trigger needed.
Layer
What to put here
System Prompt
Role: research analyst. Tone: terse - one line of “why this matters” per item. Hard rule: exactly five items, never more.
Skills
/morning-digest - fetch, rank, post the top five. /deep-dive - expand on one item when asked. /add-source - append a new feed to the source list.
Wiki
Topic taxonomy - what counts as “worth our attention.”
At 10 people someone remembers birthdays. At 50 they don’t. The agent reads a one-line-per-teammate file and posts a warm wish in #general when today matches. That’s the whole agent - no integrations, no skills, one file. Tiny ritual, outsized effect on culture. Ideal first agent because it proves the whole pattern in thirty minutes.
Schedule it: 0 10 * * * → prompt “Check if today matches any birthday in birthdays.md. If yes, post a warm message to #general. If no, do nothing.” Runs daily, silent unless it finds a match.
Layer
What to put here
System Prompt
Role: culture bot. Tone: warm, one line. Hard rule: do nothing when no birthday matches today.
Skills
None - the job is small enough to live entirely in the system prompt.
Wiki
One page, birthdays.md - name and date per teammate.
A new hire joins, gets linked to a 200-page handbook, reads 10% of it, asks the same questions every Monday for a month. Every previous hire asked them too. The agent lives in #new-hires and answers from the handbook wiki. When it doesn’t know, it names the human who does. The gaps in the handbook become visible - “this is the third ask about expense policy and the handbook doesn’t cover it.”
Layer
What to put here
System Prompt
Role: people ops assistant. Tone: warm but direct, cite the handbook section. Hard rule: when unsure, name a human - don’t guess.
Skills
/handbook - answer a question from the wiki. /request-access - kick off Drive / Notion access grants. /first-week-checklist - post the onboarding plan.
One agent saves someone an hour a week. That’s nice.Ten agents, each owning a piece of the weekly ritual, change how the team operates. The PM no longer writes release notes. The oncall rotation is less punishing. New hires ramp in days instead of weeks. The data team does real analysis instead of ad-hoc queries. Designers and engineers speak the same language because the design system stays coherent.This is the real shift - not “we added a bot” but “we have a team of specialists that handles the routine so the humans focus on judgment calls.”A boss agent on top makes it navigable. Users @ one boss with any question, and the boss decides which specialist to delegate to. The user doesn’t need to know the org chart of agents.
