Home / Engine / Agents and Council Architecture

Affirmology: Subconscious Operating System

Updated Jun 11, 2026 · Affirmology_SubconsciousOS_AgenticArchitecture_v1.md

Affirmology: Subconscious Operating System

Agentic Architecture v1

Prepared for Jeff Parker Frame Affirmology is your Subconscious Operating System. Cosmic-blueprint details are your settings. Personalized audios are software updates. Reports are diagnostics. Pilgrim is a major OS upgrade. The community is the support forum. The agentic architecture is the kernel that keeps the whole OS healthy without you having to micromanage it. Goal of this document Define an agent architecture that creates, improves, and quality-checks Affirmology's content suite autonomously while staying accountable to humans and resistant to drift. By June 20 launch the system runs itself overnight and surfaces a morning briefing that's worth reading.

Part One: Why an Agentic System At All

Single-script automation produces consistent output. Agents produce responsive output. The difference matters because:

1. The corpus changes every day. New articles get scraped, new structured records get added. A static script generator can't take advantage of new material. An agent re-queries each cycle.
2. Quality drifts. Every time we ship a new audio variant, scripts subtly shift toward whatever pattern worked last. Without a style enforcer agent watching, "small adjustments" compound into a different product after 50 renders.
3. Use cases multiply faster than I can hand-code them. Sol's Gemini conversation alone surfaced six new use cases. The team interface will generate dozens more. We need a system where adding a use case is "write a new agent" not "rewrite the script generator."
4. Cost discipline. Without budget agents watching, LLM spend can balloon when no one's looking. With them, every agent justifies every dollar.
5. Sol and Colin should not have to learn the codebase to test ideas. They type a vision into Claude, Claude dispatches the right agent, results come back. That requires an agent layer.

The orchestrator (v3) already proves the pattern works. v4 generalizes it.

Part Two: The Three-Tier Architecture

                ┌──────────────────────────────────────┐
                │  TIER 1: CONDUCTOR (meta-agent)      │
                │  - Schedules domain orchestrators     │
                │  - Allocates time and cost budgets   │
                │  - Writes the morning briefing       │
                └────────────────┬─────────────────────┘
                                 │
            ┌────────────────────┼─────────────────────┬─────────────────────┐
            │                    │                     │                     │
   ┌────────▼─────────┐ ┌────────▼─────────┐ ┌─────────▼────────┐ ┌─────────▼────────┐
   │ TIER 2: CORPUS   │ │ TIER 2: AUDIO    │ │ TIER 2: PRODUCT  │ │ TIER 2: QUALITY  │
   │ ORCHESTRATOR     │ │ ORCHESTRATOR     │ │ ORCHESTRATOR     │ │ ORCHESTRATOR     │
   │ (already v3)     │ │ (build for v4)   │ │ (build for v5)   │ │ (build for v6)   │
   └────────┬─────────┘ └────────┬─────────┘ └─────────┬────────┘ └─────────┬────────┘
            │                    │                     │                     │
            ▼                    ▼                     ▼                     ▼
   ┌─────────────────────────────────────────────────────────────────────────────────┐
   │ TIER 3: SPECIALIZED AGENTS (the working layer)                                  │
   │                                                                                 │
   │  Corpus: Scraper, Discovery, Structurer, Re-extractor                           │
   │  Audio:  Script-gen specialists (one per format), Voice-synth, Mixer, Audio QC  │
   │  Product: Assessment, Journey-design, Personalization, Companion-text           │
   │  Quality: Style-enforcer, Chart-claim verifier, Drift-detector, Cost-monitor    │
   │  Meta:   Researcher, Self-improver, Reporter                                    │
   └─────────────────────────────────────────────────────────────────────────────────┘

Communication pattern. Agents communicate through shared database state, not direct function calls. Each agent leaves findings (observations) or proposals (suggested actions) in tables. The Conductor or relevant Domain Orchestrator selects which proposals to execute based on priority + budget + recency. This pattern means new agents can be added without modifying existing ones - they just start writing to the shared tables.

Part Three: The Agent Inventory

Corpus Orchestrator agents (v3, shipped)

Audio Orchestrator agents (v4, build next)

Product Orchestrator agents (v5)

Quality Orchestrator agents (v6)

Meta agents (always running)

Part Four: How a User Request Flows Through

Concrete example. Sol opens Claude in her shared Project and types:

"Build me a 12-minute audio about navigating my breakthrough with my mother. Use my Cancer Moon and Taurus North Node anchors. Combine astrology with Gene Keys 27.4. Don't use heavy meditation music - go ambient. Companion PDF that names the three phases my mother will go through."

Inside the system:

1. Sol's Claude (in the Project) receives the request. It has tool access to the orchestrator's public API.
2. Conductor routes it to the Product Orchestrator.
3. AssessmentAgent confirms Sol's chart is on file. It has been since she became a tester. It pulls her natal data.
4. JourneyDesignAgent reads her chart + the request. It identifies: format = CrisisCompanion (not standard sankalpa, because of "breakthrough with mother" framing). Length = 12 min. Music = ambient pad, not theta meditation. Companion text needed: "three phases" structure.
5. PersonalizationAgent queries the corpus for: Cancer Moon interpretive language, Taurus North Node language, Gene Key 27.4 in independent voice (Ashley Mosaic blog material), mother-daughter dynamics from the family-dynamics literature. Builds a context bundle.
6. CrisisCompanionSpecialist (under Audio Orchestrator) receives the context bundle + Sol's chart + the request parameters. Generates the 12-min script.
7. StyleEnforcerAgent scans the draft. Catches one em dash. Returns to specialist for revision.
8. ChartClaimVerifierAgent scans the revised draft. All claims trace to Sol's chart JSON. Approved.
9. VoiceSynthesizerAgent calls ElevenLabs. Voice = Charlotte at speed 1.07. ~$0.50.
10. AudioMixerAgent runs the six-stage FFmpeg chain. Music = ambient pad from library, not theta meditation track. ~10 sec.
11. AudioQCAgent scans the final mp3. No dropouts. LUFS = -16.2. Pass.
12. CompanionTextAgent generates the PDF. Names mother's three phases (tactical volley → victim flip → background recruitment) per the family-dynamics research. Includes Sol's chart-specific anchors.
13. DeliveryAgent packages mp3 + PDF + transcript. Drops them into Sol's shared folder.
14. Sol's Claude responds in the original chat with the file links and a one-paragraph summary of what was made and why.

Total wall-clock time: about 3-4 minutes. Total cost: about $0.80. Sol's effort: one paragraph of typing.

If the system thinks the request is unusual or risky (e.g., a new format we haven't shipped before), the Conductor escalates: "I can build this, but it's a new format - confirm pricing and proceed?" Sol confirms or revises. Human stays in the loop on edge cases.

Part Five: Drift Prevention

The biggest risk of an autonomous agent fleet is silent drift - every cycle the system gets 1% more pop-spiritual, 1% less precise, 1% more wordy, until 50 cycles later the product is unrecognizable.

Defenses:

1. Baseline gold-standard renders. The original Jeff v5 script is the canonical brand voice. Sol's Gemini conversation outputs are the canonical depth target. These are stored as immutable references. The StyleEnforcerAgent compares each shipped script against the baseline embedding distance.

2. Style rules as code. Em dashes, second-person violations, hallucinated placements, trademarked terms (Joe Dispenza, ThetaHealing if not licensed), brand-voice red-flag phrases - all encoded as deterministic checks, not LLM judgment.

3. Drift score per agent. Every agent's outputs get embedded. The DriftDetectorAgent compares the rolling 7-day average against the baseline. Drift score > threshold triggers a flag to the human review queue.

4. Periodic regression testing. Every Monday, the AccuracyAuditorAgent re-runs known fixtures (Jeff v5, Josh Parini, Sol's chart). If any chart claim is wrong, system pauses and alerts.

5. Human review queue. A small sample (~3% of shipped output) gets pulled for Jeff or Sol to spot-check via the team interface. Their thumbs-up or thumbs-down feeds the SelfImproverAgent.

6. Prompt custody. No agent can modify another agent's system prompt without going through the PromptCustodianAgent, which logs every change with rationale and rollback capability.

7. Cost gates. Any agent that exceeds 2x its rolling-average daily cost gets paused and flagged.

Part Six: Accountability Mechanisms

Each agent is accountable to one metric and one orchestrator:

Weekly, the Conductor reviews each agent's metric. Underperformers get paused or their system prompt gets sent to PromptCustodian for review. Persistent underperformers get retired.

This creates real evolutionary pressure: agents that produce value continue; agents that don't get sunset. Jeff and Sol never have to manually decide "is this agent still useful?" - the metrics decide.

Part Seven: Sol's Interface (How Humans Actually Use This)

Three patterns, simplest first:

Pattern 1 (build now, weekend): "Cowork" style

Sol opens Cowork on her laptop. The shared ~/affirmology-shared/ folder is mounted via SSHFS from the Mac mini. Sol types into Cowork. Cowork reads her chart from the shared folder, calls the orchestrator's CLI via SSH, gets back files in the shared folder, presents them as computer:// links.

Pros: ships in 2 days. Same tooling pattern Jeff uses today. Cons: requires Cowork on Sol's machine, requires SSH literacy if anything breaks.

Pattern 2 (build in two weeks): "Claude.ai Project" style

Anthropic Team plan with Jeff, Sol, Colin. Shared "Affirmology" Project with pinned documents (charts, system prompts, research docs). Sol opens claude.ai in her browser, opens the Project, types her request. The Project has tool access to the orchestrator's HTTP API (exposed by the Mac mini via Tailscale). Claude calls tools, returns links.

Pros: zero install for Sol. Works from her phone. Native chat experience. Cons: requires Anthropic Team subscription ($25-30/user/mo), requires the HTTP API wrapper on the Mac mini (2-3 days of work), requires Tailscale on Sol's devices.

Pattern 3 (build in a month): "demo.affirmology.ai" style

The demo.affirmology.ai page on Cloudflare exposes a form. The form posts to the Mac mini's HTTP API. The API returns delivery URLs. Sol uses the same page anyone else would.

Pros: same surface customers use. Tests the customer flow. Cons: form is less conversational than Claude chat. Iteration on requests is harder.

My recommendation: Pattern 1 this weekend (gets Sol playing). Pattern 2 by end of next week (when Mac mini is up). Pattern 3 by July 1 (customer-facing demo). All three coexist long-term - different users prefer different surfaces.

Part Eight: What Ships When

Three-week implementation roadmap.

Week of June 11-15 (this week). - Mac mini setup (you, when HDMI splitter arrives) - Tier 2 Audio Orchestrator skeleton: AudioOrchestrator, ScriptSpecialistFleet base class, VoiceSynthesizerAgent, AudioMixerAgent, AudioQCAgent. Refactor of current agents/script_generator.py into specialist pattern. - Tier 4 Quality agents that are easiest: StyleEnforcerAgent (deterministic), ChartClaimVerifierAgent (formalize existing verifier.py). - Audio mastering upgrades from the production research doc: pedalboard reverb pass on voice, sidechain ducking, loudnorm to -16 LUFS. Ships as a flag on the existing pipeline. - Sol Pattern 1 interface (SSHFS + Cowork).

Week of June 16-22 (the LLC launch week). - Three audio specialists shipped: SankalpaSpecialist, AstrologyWalkthroughSpecialist, EFTTappingSpecialist. - AssessmentAgent v1 (the 5-min structured conversation, simplest possible version). - JourneyDesignAgent v1. - ResearcherAgent: monitors corpus growth, surfaces new candidate use cases. - DriftDetectorAgent v1 (embedding-distance baseline).

Week of June 23-29. - Anthropic Team plan + Pattern 2 interface for Sol/Colin. - GeneKeysJourneySpecialist, HumanDesignWalkthroughSpecialist, HeartCoherenceJourneySpecialist. - CompanionTextAgent generating PDF workbooks. - CheckInAgent skeleton (manual approval before sending, for now). - MorningReporterAgent upgraded to synthesize all orchestrators, not just corpus.

Week of June 30 - July 6. - Pattern 3 interface wired to demo.affirmology.ai. - All remaining specialists. - Full SelfImproverAgent loop running weekly review. - Cost-budget agents enforcing tier ceilings.

By July 6 the system runs itself overnight, you and Sol get a morning briefing, and Sol can spin up new audio formats by talking to Claude.

Part Nine: Use Cases You Can Test This Week With No New Audio

While Audio Orchestrator gets built, here are six use cases that need only existing scripts + ElevenLabs + Claude:

1. Assessment + curated playlist. User answers 5 questions, gets 3 existing audios assigned. No new audio production. Tests the personalization flow.
2. Companion PDF reports. For each existing audio, generate a companion written piece. Pure Claude. Shippable today.
3. Daily transit email. Read user's chart, compute today's transits, write a 2-paragraph reading. No audio needed. Daily habit loop.
4. Family dynamics text report. Two charts in, the deep family-dynamics narrative out (mom's Sun sign, dad's Sun sign, sibling's Sun sign overlaid on user's chart). Pure text. Sol's Gemini conversation proves the demand.
5. Synastry text report. Two charts in, the relationship/business compatibility narrative out. Pure text first; audio variant later.
6. Astrocartography report. One chart in, "where on Earth does each planetary line fall, what does that mean for your career/love/health" out. Pure text + map images. Cheap and unique.

These six ship in a week with zero new audio production work. Each becomes a test format the team interface can validate before we commit ElevenLabs spend.

Part Ten: Revised Pilgrim Pricing

Your push back: $497 is too much without an in-person component. Agreed. The revised tier:

Pilgrim Audio Journey: $297. - Includes: 5-min assessment, 90-min hyper-personalized audio across two listening sessions, companion PDF workbook, 30-day daily check-in audios (each 3-5 min, transit-timed) - Cost to deliver: ~$8 in ElevenLabs and Claude spend - Margin: 97% - Comparable to: written-only Erin Claire Jones Blueprint book ($95) but with 90 min of audio added - Positioning: "Liz Greene's Astro*Intelligence reborn for the AI era"

Inner Circle Cohort with monthly group call: $97/mo. - Includes: all Inner Circle features + monthly 60-min group call with Sol (or eventually hired facilitator) - Cost to deliver: 60 minutes of Sol's time monthly + Zoom - Margin: 95% at 50 members - Positioning: "the community where your Subconscious OS gets upgrades together"

Live Retreat Pilgrimage (future, post-launch): $1,500-3,000. - 3-day in-person retreat, 12-20 attendees, Sol or licensed facilitator - Includes the Pilgrim Audio Journey as a prerequisite - Positioning: "your OS goes from upgrade to total rebuild, in a room with your people"

This way Pilgrim Audio Journey is the automated high-value entry, Inner Circle is the recurring revenue with light human touch, Live Retreat is the high-ticket transformation with full community immersion.

Part Eleven: What I Need From You To Proceed

Three decisions and one greenlight.

1. Greenlight the Week-of-June-11-15 implementation list (Audio Orchestrator skeleton, audio mastering upgrades, Sol's Pattern 1 interface). If yes, I start tomorrow morning while the corpus orchestrator continues its work.
2. Pick which two audio specialists to ship first. I'd recommend SankalpaSpecialist (closest to current pipeline, fast win) and AstrologyWalkthroughSpecialist (highest-novelty, validates the "this is my Virgo Sun in 8th house" affirmation format). But your call.
3. Confirm the Pilgrim pricing revision ($297 audio-only / $97/mo Inner Circle / future $1500-3000 live retreat). Or push back with what feels right.
4. Pick which two of the six no-audio-needed use cases I prototype this week as text products. My recommendation: family dynamics text report (Sol's Gemini conversation proves demand) and synastry text report (your business partnership with Sol is a built-in test case).

Once those four answers come in, I work through the Week-of-June-11-15 list autonomously while you and Sol get Mac mini set up and start playing with the Pattern 1 interface.

End of v1. This is the architecture. Push back, revise, or greenlight.