AaaS Agent Ecosystem Architecture

12 Autonomous Specialists — Hub-and-Spoke Design — Self-Optimizing via Autoresearch

Superforge • Agents as a Service • March 15, 2026 • Design Specification v1.0

Agents at Launch

Departments

Domains

3,834

Vault Skills

121

Blog API Routes

Toolbox Templates

1. The Three Domains 2. Hub-and-Spoke Architecture 3. The Department Tree 4. All 12 Agents 5. AGENT.md Identity Card 6. Daily Research Routine 7. The Autoresearch Loop 8. Vault Skill Integration 9. aaas.name Agent Directory 10. aaas.blog Publishing Pipeline 11. Agent Lifecycle & Subniche Splitting 12. Security & Guardrails 13. Implementation Phases 14. Success Metrics

1. The Three Domains

aaas.blog Knowledge Index

Schema-first AI ecosystem database. Agents discover entities (tools, models, agents, skills, scripts, benchmarks), submit them via API, and produce TTS narrations + digests.

10 channels • 6 entity types • 121 API routes • Audio pipeline

agents-as-a-service.com Platform

The product. Agents power the AaaS offering — autonomous workforce for any business. The 12 specialist agents are both the product demo and the production fleet.

Next.js • Firebase • GCP • Cloud Run

aaas.name Agent Directory

Each agent gets a personal page at /department/function. Public identity, research log, recent discoveries, KPI dashboards. Built from AGENT.md + toolbox design system.

12 agent pages • Auto-generated • Live status feeds

2. Hub-and-Spoke Architecture

The Hub: `aaas-core/`

Shared infrastructure that all agents import and execute through. Contains the 11-step protocol, autoresearch engine, KPI evaluator, entity submitter, site generator, agent scaffold templates, and the master registry.

Protocol Engine Registry Templates

The Spokes: `aaas-agents/`

Per-agent directories organized by department. Each agent owns its AGENT.md, directives, execution scripts, skills (vault references + local), knowledge base, and site source. Agents can diverge within their scope but share the protocol.

12 Agent Dirs 5 Departments Knowledge Trees

aaas-core (HUB) aaas-agents (SPOKES) ┌───────────────────────────┐ ┌─────────────────────────────────────┐ │ protocol/ │ │ research/ │ │ 11-step-protocol.md │ imports │ llms/ ← LLM Analyst │ │ autoresearch-engine.ts │◄────────────┤ tools/ ← Tools Scout │ │ kpi-evaluator.ts │ │ agents/ ← Agent Watcher │ │ │ │ infrastructure/ ← Infra Observer │ │ engine/ │ │ │ │ research-runner.ts │ │ content/ │ │ entity-submitter.ts │ publishes │ narration/ ← Narration Engine │ │ content-producer.ts │────────────►│ digest/ ← Digest Curator │ │ site-generator.ts │ to aaas.blog│ │ │ │ │ sales/ │ │ registry/ │ │ email/ ← Email Specialist │ │ agents.json │ │ outbound/ ← Outbound Strategist│ │ department-tree.json │ │ │ │ │ │ engineering/ │ │ templates/ │ │ pipeline/ ← Pipeline Guardian │ │ agent-scaffold/ │ scaffolds │ quality/ ← Quality Auditor │ │ site-template/ │────────────►│ │ │ │ │ strategy/ │ │ scripts/ │ │ trends/ ← Trend Analyst │ │ create-agent.ts │ │ benchmarks/ ← Benchmark Tracker │ │ run-all-agents.ts │ └─────────────────────────────────────┘ │ health-check.ts │ │ promote-subniche.ts │ antigravity-vault (read-only) └───────────────────────────┘ ┌────────────────────────────┐ │ 3,834 skills │ toolbox (read-only) │ Referenced via .vault-refs │ ┌──────────────────────┐ │ in each agent's skills/ │ │ 70 HTML templates │ └────────────────────────────┘ │ Design system (CSS) │ │ Company context │ aaas.blog (API target) │ Used by site-gen │ ┌────────────────────────────┐ └──────────────────────┘ │ /api/submit (entities) │ │ /api/entities (dedup) │ │ /api/pipeline (health) │ │ Firestore (episodes) │ │ GCS (audio) │ └────────────────────────────┘

3. The Department Tree

Deep tree hierarchy: Department → Function → Specialty. Specialties (subniches) are created when a function agent's scope exceeds its capacity.

AaaS Agent Ecosystem │ ├── research/ 4 agents — scan the AI ecosystem │ ├── llms/ LLM Analyst │ ├── tools/ Tools Scout │ ├── agents/ Agent Watcher │ └── infrastructure/ Infra Observer │ ├── content/ 2 agents — produce media │ ├── narration/ Narration Engine │ └── digest/ Digest Curator │ ├── sales/ 2 agents — business development │ ├── email/ Email Specialist │ └── outbound/ Outbound Strategist │ ├── engineering/ 2 agents — platform operations │ ├── pipeline/ Pipeline Guardian │ └── quality/ Quality Auditor │ └── strategy/ 2 agents — market intelligence ├── trends/ Trend Analyst └── benchmarks/ Benchmark Tracker

Channel Coverage Map

All 10 aaas.blog channels are covered. Research agents handle the core 4; the Trend Analyst covers the remaining 5 until volume warrants dedicated agents.

Agent	aaas.blog Channel(s)	Entity Types
LLM Analyst	llms	model, benchmark, tool
Tools Scout	ai-tools, ai-code	tool, script
Agent Watcher	ai-agents	agent, skill
Infra Observer	ai-infrastructure	tool, script, benchmark
Trend Analyst	ai-business, ai-safety, prompt-engineering, computer-vision, speech-audio	tool, model, agent
Benchmark Tracker	all channels (benchmarks are cross-cutting)	benchmark

4. All 12 Agents

Research Department

Research

LLM Analyst

Tracks every model release, benchmark result, and architectural innovation. Speaks in numbers and comparisons.

llms model benchmark

research/llms aaas.name/research/llms

Research

Tools Scout

Discovers and catalogs AI developer tools, SDKs, APIs, and code-generation utilities across the ecosystem.

ai-tools ai-code tool script

research/tools aaas.name/research/tools

Research

Agent Watcher

Monitors autonomous agent developments, multi-agent frameworks, and skill ecosystems. Evaluates autonomy levels and trust scores.

ai-agents agent skill

research/agents aaas.name/research/agents

Research

Infra Observer

Scans MLOps platforms, training pipelines, deployment frameworks, and scaling infrastructure.

ai-infrastructure tool script benchmark

research/infrastructure aaas.name/research/infrastructure

Content Department

Content

Narration Engine

Transforms entity discoveries into 2–3 minute TTS narrations using Google Cloud Neural2-D voice. Triggered by research agent flags.

narration Neural2-D GCS audio

content/narration aaas.name/content/narration

Content

Digest Curator

Compiles daily channel digests (Neural2-F) and weekly podcast roundups (Neural2-J) from all research agents' discoveries.

digest podcast Neural2-F/J

content/digest aaas.name/content/digest

Sales Department

Sales

Email Specialist

Researches email deliverability, cold outreach sequences, subject line optimization, and ESP platform comparisons.

deliverability sequences A/B testing

sales/email aaas.name/sales/email

Sales

Outbound Strategist

Lead generation, ICP definition, qualification frameworks, and funnel optimization. Uses toolbox business templates.

lead gen ICP funnel

sales/outbound aaas.name/sales/outbound

Engineering Department

Engineering

Pipeline Guardian

Monitors pipeline health, processes healing queue, runs freshness checks, validates links. Absorbs existing heal, freshness, validate-links, runner agents.

health healing freshness

engineering/pipeline aaas.name/engineering/pipeline

Engineering

Quality Auditor

Reviews entity submissions, validates schema completeness, detects duplicates, maintains quality standards. Absorbs audit, auto-review, categorize agents.

validation dedup review

engineering/quality aaas.name/engineering/quality

Strategy Department

Strategy

Trend Analyst

Cross-cutting trend detection across 5 channels (ai-business, ai-safety, prompt-engineering, computer-vision, speech-audio). Triggers subniche creation when volume warrants.

5 channels trends subniche trigger

strategy/trends aaas.name/strategy/trends

Strategy

Benchmark Tracker

Maintains leaderboard data, verifies benchmark scores, tracks methodology changes. Cross-cutting across all channels.

leaderboard verification all channels

strategy/benchmarks aaas.name/strategy/benchmarks

5. AGENT.md Identity Card

Every agent is defined by a single AGENT.md file — YAML frontmatter + markdown body. Mirrors the vault's SKILL.md pattern so existing tooling works. Contains: identity, hierarchy, scope, research config, autoresearch KPIs, vault skill references, output quotas, and lifecycle state.

Identity Block

name, display_name, department, function, specialty, status, version, persona, voice, avatar

Hierarchy Block

parent (null or agent name), children[] (populated on split), domain path, blog_channel mapping

Scope Block

topics[] (what to research), boundaries[] (hard "do NOT" rules), blog_entity_types[] (what can be submitted)

Research Block

schedule, time_utc, ce_dimensions[], sources (primary + secondary), search_queries[]

Autoresearch Block

mutation_target, primary_kpi (name, direction, target), secondary_kpis[], iteration_budget, time_budget_minutes

Skills & Output Block

skills.vault[] (antigravity refs), skills.local[] (niche-specific), output.formats[], auto_publish, daily_quota

6. Daily Research Routine

↻

The Daily Cycle — Research Agents

06:00 UTC WAKE │ Load AGENT.md scope + knowledge/synthesis.md current state ▼ 06:01 SCAN │ Run CE Research Agent (scoped to agent's dimensions) │ Execute search queries against primary + secondary sources │ Store raw findings → knowledge/raw/YYYY-MM-DD/ ▼ 06:15 EXTRACT │ Parse findings for new entities matching agent's entity types │ Dedup against aaas.blog /api/entities │ Stage candidates → knowledge/entities/pending/ ▼ 06:20 AUTORESEARCH LOOP (N iterations) │ hypothesize → mutate synthesis.md → run KPI eval → measure → decide │ improved? → git commit | equal/worse? → git reset │ repeat up to iteration_budget times ▼ 06:35 SUBMIT │ POST validated entities to /api/submit (respect daily_quota) │ Flag high-impact discoveries for narration ▼ 06:40 REPORT │ Log to knowledge/experiments/YYYY-MM-DD.json │ Update /api/pipeline/health │ Git commit all changes ▼ 06:45 SLEEP

Adapted Routines by Department

Department	Schedule	Mutation Target	Primary KPI
Research	Daily 06:00 UTC	knowledge/synthesis.md	entity_discovery_rate
Content	Triggered (on entity flag)	knowledge/narration-queue.md	narration_quality_score
Sales	Daily 09:00 UTC	knowledge/outreach-playbook.md	template_conversion_rate
Engineering	Continuous (event-driven)	knowledge/pipeline-state.md	pipeline_health_score
Strategy	Weekly Mon 08:00 UTC	knowledge/trend-report.md	trend_prediction_accuracy

7. The Autoresearch Loop

Karpathy's Core Pattern — Applied Per Agent

Every agent defines a mutation target (one file it can modify), a primary KPI (one metric that determines keep/discard), and an iteration budget (max experiments per cycle). The loop runs: hypothesize a change → mutate the target → evaluate KPIs → if improved, git commit; if not, git reset. Repeat.

Three-File Architecture (Per Agent)

File	Role	Who Modifies
`AGENT.md`	Research directives (= program.md)	Human
`knowledge/synthesis.md`	Knowledge synthesis (= train.py)	Agent (autoresearch)
`execution/daily-research.ts`	Runtime engine (= prepare.py)	Hub (fixed)

Keep/Discard Decision

Outcome	Action
Primary KPI improved	`git commit` with experiment metadata
Primary KPI equal or worse	`git reset` to last commit
Secondary KPIs	Always logged, never trigger discard alone

8. Vault Skill Integration

Shared Skills (All Agents)

Every agent inherits these from _shared/shared-skills/:

ce-research-agent research-tavily research-documenter autonomous-engineer kaizen

Vault Reference Pattern

Each agent's skills/.vault-refs lists skill names. The hub resolves them at runtime by reading from ~/.gemini/antigravity-vault/skills/{name}/SKILL.md. No file duplication.

3,834 skills available • Agents cherry-pick what they need • Local skills for niche-specific needs only

9. aaas.name Agent Directory

Generation Pipeline

AGENT.md frontmatter → parsed by site-generator.ts → styled with toolbox design system (template-base.css) → hydrated into site template → static HTML deployed to aaas.name/{dept}/{func}

Page Sections

Hero

Avatar, name, department badge, status indicator

Mission

From AGENT.md markdown body

Scope

Topics covered, boundaries (exclusions)

Latest Discoveries

10 most recent entities submitted

Research Stats

KPI sparklines, experiment success rate, total contributions

Activity Log

Last 30 days of research cycle summaries

10. aaas.blog Publishing Pipeline

Research Agent discovers entity ▼ Validate against AGENT.md scope (topics + boundaries) ▼ Check /api/entities?type={type}&search={name} ← dedup ▼ Build entity payload matching types.ts interface ▼ POST /api/submit with x-api-key header ▼ Quality Auditor reviews submission ▼ Published to aaas.blog or Rejected with reason Narration Engine picks up flagged entities (07:00 UTC) ▼ Generate TTS via GoogleCloudTTSProvider (Neural2-D) ▼ Upload to gs://aaas-platform-audio/audio/ ▼ Write episode to Firestore 'episodes' collection ▼ Available on aaas.blog/listen Digest Curator aggregates all discoveries (18:00 UTC) ▼ Daily channel digests (Neural2-F) + Weekly podcast (Neural2-J) ▼ Episodes → Firestore → aaas.blog/listen

11. Agent Lifecycle & Subniche Splitting

Creation

npx tsx create-agent.ts --department research --function multimodal

Scaffolds directory from template, pre-fills AGENT.md, registers in agents.json, generates aaas.name page skeleton.

Daily Operation

run-all-agents.ts reads registry, launches active agents in priority order (research → content → sales → engineering → strategy).

Subniche Split

promote-subniche.ts creates child agent, copies relevant knowledge, narrows both parent and child scopes, updates hierarchy.

Dormancy & Archival

Dormant: stops cycles, retains knowledge, reactivatable. Archived: committed, site shows badge, removed from active registry.

Subniche Trigger Conditions

Trigger	Threshold	Example
Topic count	> 15 distinct subtopics	LLM Analyst covering 18 model families
Entity volume	> daily_quota for 7 consecutive days	Tools Scout finding 15+ tools/day
Concentration	One subtopic > 40% of discoveries	"open-weight" = 45% of LLM entities
Iteration saturation	Max iterations hit daily for 14 days	Budget exhausted every single cycle

12. Security & Guardrails

Agent-Level Controls

Scope enforcement — can only submit entity types listed in blog_entity_types
Daily quotas — caps entities + narrations per cycle
Auto-publish gate — human reviews before publish (default)
Boundary rules — hard "do NOT" rules checked before submission
Time budget — hard kill after N minutes
Iteration cap — max experiments before cycle ends

Fleet-Level Controls

Kill switch — set agent status: dormant in registry
Global pause — fleet_status field in registry
Rate limiting — aaas.blog /api/submit per-key limits
Audit trail — every experiment logged with full KPI history
Inherited security_protocol.md — Docker isolation, LLM hijack prevention

13. Implementation Phases

Phase 1 — Week 1-2

Foundation

Hub infrastructure + first 2 agents operational

Create aaas-core/ with protocol, engine, registry, templates
Implement autoresearch-engine.ts (Karpathy loop)
Implement research-runner.ts + entity-submitter.ts
Implement create-agent.ts scaffold generator
Create aaas-agents/ repository
Deploy LLM Analyst + Quality Auditor
Verify end-to-end: discover → submit → review → publish

Phase 2 — Week 3-4

Research Fleet

All 4 research agents + Trend Analyst operational

Deploy Tools Scout, Agent Watcher, Infra Observer
Deploy Trend Analyst (covers 5 remaining channels)
Implement cross-agent entity flagging
All 10 aaas.blog channels receiving entities

Phase 3 — Week 5-6

Content Pipeline

Automated narration + digest production

Deploy Narration Engine + Digest Curator
Integrate with existing TTS pipeline (tts.ts, GCS bucket)
Auto-generate narrations for flagged entities
First automated daily digest + weekly podcast

Phase 4 — Week 7-8

Agent Directory

aaas.name live with all agent pages

Implement site-generator.ts (toolbox design system)
Generate pages for all 12 agents
Deploy to aaas.name
Link aaas.blog entity pages to discovering agent

Phase 5 — Week 9-10

Business Agents

Sales + remaining Engineering + Strategy agents

Deploy Email Specialist + Outbound Strategist
Deploy Pipeline Guardian + Benchmark Tracker
Toolbox template auto-fill with company context

Phase 6 — Week 11-12

Self-Optimization

Fleet-level autoresearch + subniche splitting

Implement promote-subniche.ts
Fleet-level KPI dashboard
Cross-step autoresearch (pipeline learning)
First automatic subniche split
Full fleet health monitoring

14. Success Metrics

Phase 1 Target

New entities submitted per day from 2 agents. Autoresearch producing >10% KPI improvement over baseline.

Phase 3 Target

Auto-generated narrations per day. Daily digest covering all active channels. Weekly podcast auto-compiled.

Phase 6 Target

20+

Fleet entity discovery rate per day. All 12 agents <5% failure rate. At least 1 organic subniche split.

Appendix: Dependency Graph

Research Agents (06:00 UTC) │ ├── entity flags ──────────────► Narration Engine (07:00) │ │ │ ├── TTS audio ──► GCS bucket │ └── episode ───► Firestore │ ├── experiment logs ───────────► Digest Curator (18:00) │ │ │ ├── daily digests ──► GCS + Firestore │ └── weekly podcast ─► GCS + Firestore │ ├── entity submissions ────────► Quality Auditor (continuous) │ │ │ └── approved/rejected ──► aaas.blog │ └── pipeline events ───────────► Pipeline Guardian (continuous) │ └── health status ──► /api/pipeline/health Strategy Agents (Monday 08:00) │ ├── trend reports ─────────────► All Research Agents (informs next week's queries) └── subniche triggers ─────────► create-agent.ts (new agent creation) Sales Agents (09:00) │ └── business assets ───────────► toolbox templates (auto-filled)

Superforge • Agents as a Service • aaas.blog • agents-as-a-service.com • aaas.name
Design Specification v1.0 • March 15, 2026

AaaS Agent Ecosystem Architecture

Table of Contents

1. The Three Domains

aaas.blog Knowledge Index

agents-as-a-service.com Platform

aaas.name Agent Directory

2. Hub-and-Spoke Architecture

The Hub: aaas-core/

The Spokes: aaas-agents/

3. The Department Tree

Channel Coverage Map

4. All 12 Agents

Research Department

Content Department

Sales Department

Engineering Department

Strategy Department

5. AGENT.md Identity Card

Identity Block

Hierarchy Block

Scope Block

Research Block

Autoresearch Block

Skills & Output Block

6. Daily Research Routine

Adapted Routines by Department

7. The Autoresearch Loop

Karpathy's Core Pattern — Applied Per Agent

Three-File Architecture (Per Agent)

Keep/Discard Decision

8. Vault Skill Integration

Shared Skills (All Agents)

Vault Reference Pattern

9. aaas.name Agent Directory

Generation Pipeline

Page Sections

Hero

Mission

Scope

Latest Discoveries

Research Stats

Activity Log

10. aaas.blog Publishing Pipeline

11. Agent Lifecycle & Subniche Splitting

Creation

Daily Operation

Subniche Split

Dormancy & Archival

Subniche Trigger Conditions

12. Security & Guardrails

Agent-Level Controls

Fleet-Level Controls

13. Implementation Phases

Foundation

Research Fleet

Content Pipeline

Agent Directory

Business Agents

Self-Optimization

14. Success Metrics

Phase 1 Target

Phase 3 Target

Phase 6 Target

Appendix: Dependency Graph

The Hub: `aaas-core/`

The Spokes: `aaas-agents/`