Bloom
Pilot · Function: Cross-functional · every department · Where it runs: Cloud Run + Vertex AI + Firestore + GCS + Google Workspace
Scotts's Agent Fleet platform — purpose-built AI agents across every department, with their output landing back in Google Workspace where the work already happens.
Problem
Different departments at Scotts need very different things from AI. Marketing needs long-form content grounded in R&D data. Finance needs reconciliation and reporting. Customer Service needs research synthesis. Merchandising needs media reporting. Each one wants AI to do the work and hand it back to them in the tools they already use — not be redirected to yet another portal to copy-paste from.
Building each agent as a one-off script on a laptop creates a fragmented, ungoverned mess. Building each one as a bespoke web app means re-inventing the wheel every time. We need one platform where:
- New agents ship by adding a skill, not standing up infrastructure
- Every agent is governed, observable, and cost-controlled by default
- Output lands in Workspace — Docs, Sheets, Drive, Gmail — so it shows up where work already lives
What Bloom is
A managed fleet of agents, one platform, output back into Workspace.
- One conversation per session — each user session runs as a Cloud Run Job execution wrapping the Claude Agent SDK + Claude CLI, with one Claude conversation context per session for warm cache reuse across turns.
- Skill library — versioned in a GCS bucket, synced into each session at boot. Adding a new agent (or extending one) means adding a skill directory with
SKILL.md+ tools + agent prompts. No platform changes, no new deployments. - Two-tier model routing — Claude Haiku orchestrates and dispatches tools; specialist subagents run heavier models (Claude Opus for long-form composition, etc.). Subagent isolation keeps the orchestrator's context clean. ~50–70% cost reduction vs. all-Opus per session.
- Workspace as the output surface — finished work lands directly in Google Drive (shared, scoped folders per agent/department), Docs are edited in place, comments get replied to. Users stay in the apps they already live in.
- Context hygiene built in —
AGENTS.mdlists tools + subagents per turn; scratchpad pattern keeps heavy tool outputs out of the conversation; proactive compaction at 70% context usage. - Governed by default — Google IAP fronts the platform with a
scotts.comdomain allowlist; no per-user provisioning. Every session has cost, context, and event telemetry in Firestore. Every tool call is auditable.
The fleet today
Each agent is a thin orchestrator that interacts with the user, then fires a corresponding worker (Cloud Run Job) for the heavy lifting — LLM calls, tool use, Drive uploads, Apify scrapes. The orchestrators are what associates think of as "the agents"; the workers are runtime plumbing.
| Orchestrator | Department | What it does |
|---|---|---|
| My Agents (meta-orchestrator) | Cross-functional | Personal assistant of assistants — delegates to the right specialist agent and (optionally) chains follow-ups |
| Article Rewriter | Marketing · Content | SMG long-form article rewrites grounded in R&D data + Universal Search; handles user comments and edits post-publish |
| CA-EN Rewriter | Marketing · Content (Canada) | US → English-Canada article localization, 12 sections of localization rules |
| CA-FR Rewriter | Marketing · Content (Canada FR) | English-Canada → French-Canada (Quebecois) translation and localization |
| PLP Writer | Marketing · Product | Product landing-page composition with creative + render + Drive publish |
| AMZ Media Reporting | E-commerce · Merchandising | Weekly Amazon YoY pipeline — CSV ingest → YoY analysis → exec summary .docx → per-week Drive folder |
| Email Extraction | Master Data / Operations | Parses incoming vendor master-data update emails into structured JSON updates |
| Social Listening | Marketing · Brand | Goal-aware social listening sweeps via Apify, with writer subagent producing the report |
| Live Agent | Cross-functional | "Thought Partner" — persistent conversation across hours/days, holds workspace state, native HITL |
New orchestrators land as the intake backlog and SuperIntelligent sweeps surface them.
Who owns it
- Business owner: AI CoE
- Tech stack: Cloud Run Service (FastAPI), Cloud Run Jobs (worker, Claude CLI), Firestore (state), GCS (skill library), Vertex AI (Claude global endpoint), Google Drive + Workspace (publish surface), Universal Search + Discovery Engine + Gemini (external tools)
- Source:
LongFormApp(bloom-api,bloom-agents,bloom-ui) - GCP project:
mkt-dev-0-0, regionus-east5
Data it touches
| Data source | Classification | Access path |
|---|---|---|
| Product catalog (Universal Search) | Internal | Read-only via API |
| Internal R&D data (Discovery Engine) | Confidential | Read-only, IAP-authenticated |
| Web search grounding | Public | Gemini-grounded |
| Generated content | Internal | Published to scoped Google Drive folders per agent/department |
| Session state | Internal | Firestore — user email from IAP header |
Governance status
- AI Governance Committee reviewed: Internal platform, IAP-gated, no consumer-facing output today
- Human-in-the-loop required for: All generated content reviewed before external publication
- Vendor / model risk noted: Anthropic Claude via Vertex AI (enterprise data terms); Gemini for grounding + image generation
From prototype to production
Bloom is the path agentic solutions take once they outgrow a single laptop — from a Cowork prototype to a governed, observable, scalable agent. Skills that prove out land here, get versioned in the shared skill library, and become reusable across the org.
This is also the platform the AI Hub itself uses for IAP-gated hosting (same Cloud Run + IAP + scotts.com allowlist pattern).
Patterns reused
- Cloud Run + IAP + domain allowlist (used by SLCP's BFF, this hub, and every future CoE-built internal tool)
- Skill library on GCS — versioned, scoped, synced per session
- Two-tier model routing (Haiku orchestrator + Opus writer subagent) for cost efficiency on long sessions
- Workspace as the output surface — agents publish where users already work
What's next
- More agents as pilots prove out — driven by intake + SuperIntelligent sweep priorities
- MCP-server access controls for agent-to-agent communication (FY27 agentic governance)
- Champion-network onboarding so function AI Solution Leads can ship skills end-to-end
- Migration from Cowork-prototype-to-Bloom-skill becomes the standard productionization path
Changelog
- 2026-05-21 — Initial entry on the hub.