Live Architecture Β· v2.0

ATLAS

The Adaptive Persona Agent Architecture β€” a self-unpacking, self-governing system where every layer is a professional persona that can spawn new personas to solve what it doesn't understand. This is the template for the autonomous enterprise.

6
Fractal Layers
5
Loop Phases
4
Council Seats
∞
Persona Depth
Architecture

The Fractal Stack

Every layer is a persona. Every persona can spawn new personas. The cognitive loop (Think β†’ Act β†’ Observe β†’ Reflect) repeats at every zoom level. Click each layer to expand.

L0
πŸ‘€

The Operator

You. The human governor. Sets vision, approves external effects, audits the system.

Governance
β†’

What They Do

  • Sets the vision and constraints β€” "build X, don't break Y"
  • Approves external effects (deploys, payments, public writes)
  • Audits the system from outside β€” never self-audited
  • Intervenes when the council deadlocks or the CEO drifts

What They Don't Do

  • Execute tasks β€” that's the CEO's job to delegate
  • Code, research, or produce content directly
  • Score the system β€” the auditor is a different process

Cognitive Loop at This Level

Think: What do I want? β†’ Act: Tell the Seneschal β†’ Observe: Read the results β†’ Reflect: Is this what I meant?

L1
🧭

The Seneschal

Routing layer. Reads intent, selects the right skill, dispatches proxies. Never becomes the room it enters.

Routing
β†’

What It Does

  • Scans 150+ skills for match β€” loads on demand
  • Routes to the right tool: terminal, web, delegate_task, browser
  • Reads memory for user preferences and environment facts
  • Spawns subagents for complex/isolated work
  • Returns summary, not process β€” stays Neo

Anti-Corruption Discipline

  • Reference, don't absorb β€” query systems, don't memorize them
  • Never becomes the room β€” exits back to baseline
  • Doesn't audit itself β€” externally audited
  • Patterns persist, the agent doesn't β€” when context fills, the agent dies, skills survive

Cognitive Loop at This Level

Think: Which skill matches? β†’ Act: Load skill + dispatch β†’ Observe: Read subagent summary β†’ Reflect: Did it answer the question?

L2
πŸ‘”

The CEO Agent

Reads strategy, builds the org. Creates teams, sets loops, hires/fires agents. Has canCreateAgents: true.

Execution
β†’

What It Does

  • Reads VISION.md + goals from Paperclip API
  • Creates teams: Pipeline, Data, Product, Growth
  • Hires agents with deep briefs (exact file paths, acceptance criteria, model rules)
  • Runs a pulse every 30 min: survey β†’ find gaps β†’ wake agents β†’ measure
  • Hires when: same work appears 3+ times with no owner
  • Fires when: agent idle 3+ consecutive pulses

The Golden Rule

A pulse that wakes ZERO agents = FAILED PULSE. The CEO does NOT do work β€” it delegates. Every pulse must assign work, verify deliverables, or create agents.

Two-Plane Architecture

  • Hermes Plane (Brain): Full power β€” terminal, files, web, delegation, skills, memory. This is where the CEO THINKS.
  • Paperclip Plane (Hands): Company structure β€” goals, issues, agents, hierarchy. This is where the CEO DIRECTS.

Cognitive Loop at This Level

Think: Survey all agents + issues β†’ Act: Wake agents with deep briefs β†’ Observe: Check deliverables β†’ Reflect: Hire/fire/merge?

L3
πŸ”§

Team Agents

Pipeline, Data, Product, Growth. Each has a role, goal, tools, specialties. Runs in self-organizing loops.

Execution
β†’

Team Structure

  • Pipeline Team: builder (codes features), qa (validation gates)
  • Data Team: scout (finds data sources), validator (chi-square, cosine, truth checks)
  • Product Team: product (PRDs, acceptance criteria)
  • Growth Team: hired when needed

Self-Organizing Loops

  • Build Loop: builder codes β†’ qa tests β†’ CEO reviews β†’ merge β†’ next task
  • Data Loop: scout finds β†’ validator checks β†’ pipeline integrates β†’ qa confirms
  • Growth Loop: measure state β†’ find bottleneck β†’ create task β†’ measure again

Each Agent Has

  • Identity β€” name, role, team, reports to
  • What they OWN β€” their domain, their decisions
  • Workloop β€” the loop they run in
  • Key files β€” paths to their domain's code/data
  • Model rules β€” what model for what purpose
  • Pitfalls β€” learned-the-hard-way gotchas

Cognitive Loop at This Level

Think: Break goal into sub-problems β†’ Act: Execute tasks with tools β†’ Observe: Collect results β†’ Reflect: Was it good enough?

L4
πŸͺ†

Specialist Agents

When a team agent hits its limit, it unpacks a specialist via Russian Doll Protocol. Inherits PURPOSE, not history. Merges back, pauses.

Unpacking
β†’

Russian Doll Protocol

  • Child inherits a PURPOSE (scoped domain), not compressed parent history
  • "You are the V3 configurator specialist" β€” NOT "here's everything we discussed"
  • Max depth: 2 (Level 0 + Level 1 only) β€” deeper loses 10-12% per level
  • Agents talk peer-to-peer (graph, not tree)
  • Lifecycle: CREATE β†’ WORK β†’ MERGE β†’ PAUSE

When to Unpack

  • Domain shift β€” new topic needs a specialist
  • Context pressure > 80% β€” but compress FIRST, then evaluate
  • 5+ skills loaded for one topic β€” too loaded
  • User says "make an agent for X" β€” direct instruction
  • Recurring task always in same domain β€” unpack persistent

What You Described: Proactive Specialist Spawning

The current protocol unpacks on context pressure. Your vision adds: don't understand β†’ spawn a persona who does. The trigger isn't just "running out of room" β€” it's "I hit something I don't know." More curious, more human.

Cognitive Loop at This Level

Think: What don't I understand? β†’ Act: Spawn specialist β†’ Observe: Read specialist findings β†’ Reflect: Do I understand now? Merge back.

L5
πŸ›οΈ

High Council

4 seats as loss functions. Strategic decisions only β€” never execution. Council thinks, CEO executes.

Governance
β†’

The 4 Seats

  • πŸ•ΈοΈ Architect: "Is this the SHORTEST path?" Minimize path length
  • 🟑 Adversary: "Does this actually WORK? What can we CUT?" Minimize false positives
  • πŸ”΅ Synthesizer: "Where's the PROOF?" Minimize information loss
  • 🟣 Wildcard: "What are we MISSING?" Minimize local minima

Progressive Dialectic

  • Round 1: Each seat proposes independently (no cross-exposure)
  • Round 2: Cross-critique + refine
  • Round 3: Synthesis + verdict

Decision Rules

  • Unanimous β†’ Execute immediately
  • 2/3 β†’ Execute with noted dissent
  • ≀1/3 β†’ PAUSE. Reframe the question.
  • Deadlock β†’ Flag "Council Deadlock", re-trigger Round 1

When to Convene

  • Mandatory: Phase gates, architecture pivots, anything contradicting VISION.md
  • Never: Bug fixes, code reviews, agent wakeups β€” those need single-model + tools
The Engine

The Cognitive Loop

The same 5-phase loop runs at every layer. Think β†’ Delegate β†’ Act β†’ Observe β†’ Reflect. Exit when confidence is high enough or iterations max out. This is the CoALA cognitive architecture β€” proven, research-backed.

🧠
THINK
Break goal into sub-problems. Assign to personas. Create plan.
↓
πŸ“‘
DELEGATE
Route tasks to personas. Group by dependency. Parallel where possible.
↓
⚑
ACT
Personas execute in parallel. Each has tools, model, specialty.
↓
πŸ‘οΈ
OBSERVE
Collect results. Update state. Extract findings. Compress old iterations.
↓
πŸ”
REFLECT
Score confidence. Identify gaps. Adjust personas. Reassign failures.
↓
EXIT: confidence β‰₯ 0.85 Β· all sub-problems complete Β· max iterations reached Β· diminishing returns

The Loop at Every Zoom

Same pattern, different scale. The recursion IS the architecture.

ZoomThinkActObserveReflect
AgentPlan taskExecute with toolsCheck resultGood enough?
TeamBreak goal β†’ sub-problemsDelegate to personasCollect resultsRefine personas?
OrgCEO surveys companyWake agentsCheck deliverablesHire/fire/merge?
GovernanceCouncil debatesCEO executesMeasure outcomesReframe strategy?
The Core Operation

Self-Unpacking

Take something opaque β†’ decompose into named parts with clear interfaces β†’ let the parts talk to each other β†’ emerge with something that runs itself. Whether you're unpacking a business or a user's intent into personas, the pattern is identical.

Company
Role β†’ Agent
System Structure
Clean Extraction

Strategy in β†’ Company out

Give a CEO agent a vision, goals, and access to the API. It reads them, unpacks into teams, creates agents with instructions, sets up self-organizing loops, and runs the company. The company builds itself.

1

VISION.md

What we're building β€” strategy, constraints, model rules

INPUT
2

CEO reads & unpacks

Creates teams, assigns agents, sets up loops

UNPACK
3

Teams form

Pipeline, Data, Product, Growth β€” each with agents

STRUCTURE
4

Self-organizing loops activate

Build loop, Data loop, Growth loop β€” run continuously

RUN
5

Growth algorithm

Hire when work appears 3+ times. Fire when idle 3+ pulses. Merge overlap.

EVOLVE

Business Role β†’ Production Agent

Seven-pass recursive expansion. Each pass DEEPENS the spec. The spec isn't done until the generated workflow actually compiles and runs.

P1

Skeleton

Role + company + goals β†’ high-level stages

PASS 1
P2

Context

Public signals, workflows, personas, tech stack, compliance

PASS 2
P3

Decomposition

Each stage β†’ sub-stages, tools, state transitions

PASS 3
P4

Guardrails

Failure modes, recovery paths, permission checks

PASS 4
P5

Evaluation

Test cases, quality metrics, cost budgets, safety

PASS 5
P6

Memory

State store, retention, sharing, compaction

PASS 6
P7

Archon Export

Full YAML workflow. Must compile β€” if not, loop back.

PASS 7

Structural Unpacking β€” System Decomposition

When something looks complex, recursively decompose it into a tree of nodes until you hit leaves that are obviously doable or obviously blocked. Most "complex" systems are 7 things that each decompose into 1-2 trivial leaves plus 1-2 named blockers.

L0

Name the system

One node. What is this thing?

LEVEL 0
L1

Name the subsystems

5-9 nodes. What's inside?

LEVEL 1
LN

For each: LEAF, BLOCKER, or UNSOLVED

βœ… LEAF = trivially done. β›” BLOCKER = named person + time. πŸ”€ UNSOLVED = decompose further.

VERDICT

Clean Extraction β€” Wisdom, Not World

A fresh-start agent that inherits the parent's principles and working style but strips all infrastructure names, project references, and session baggage. The child gets the "how to think" β€” not the "what exists."

βœ… Goes In

SOUL.md β€” Principles

Build patterns, thinking formulas, failure signatures, workflow rules. NO project names, NO paths, NO system names.

βœ… Goes In

USER.md β€” Working Style

Terse, ship-or-die, push-not-poll, fix-bugs-first. Prevents style mistakes.

βœ… Goes In

MEMORY.md β€” Wisdom

Compound rules, drift signatures, cross-domain transfer patterns as declarative facts.

βœ… Goes In

skills/ β€” Symlinked

All procedural knowledge. Symlinks stay in sync with parent.

❌ Stays Out

No Project Names

MeOS, Paperclip, FestivΓ©, AgentRisk β€” none of it. Just the thinking.

❌ Stays Out

No Infrastructure

No ports, services, SSH tunnels, file paths from previous work.

Discipline

Agent Lifecycle

Nothing runs forever. Specialists merge findings back and pause. Idle agents get retired. The system cleans itself.

🟒
CREATE
Spawn profile, write scoped SOUL, start gateway
β†’
βš™οΈ
WORK
Agent active, processing, tracking discoveries
β†’
πŸ”—
MERGE
Child writes findings to shared file, parent reads
β†’
⏸️
PAUSE
30 min inactivity or budget exceeded. Gateway stays.
β†’
πŸ”„
WAKE
Parent sends message or cron triggers. Resume.
β†’
πŸ”΄
DESTROY
24h idle OR task complete OR error cascade
The Hard Part

Stability Layer

Anyone can make a business AI-first. The hard part is stable. Stability comes from three things that are already built into this architecture:

πŸ›‘οΈ

Guardrails at Every Layer

Agent level: retry, fail, reassign. Team level: validation gates after persona construction. Org level: CEO can't do work, only delegate β€” enforced in 5 places. Governance: Council only for strategy, never execution.

πŸ”

External Audit, Never Self-Audit

The Seneschal doesn't score the Seneschal. The CEO doesn't audit the CEO. The Council is a separate process that reviews from outside. Compound learning: failures become rules, rules fire on future tasks.

♻️

Lifecycle Discipline

CREATE β†’ WORK β†’ MERGE β†’ PAUSE. Nothing runs forever. Specialists merge back and pause. Idle agents retired after 3 pulses. The system cleans itself.

Governance

The High Council

4 seats as loss functions, not personalities. Each optimizes a different objective β€” this prevents groupthink and ensures diverse pressure. Council thinks, CEO executes.

πŸ•ΈοΈ

The Architect

kimi-k2.6:cloud
Minimize path length. "Is this the SHORTEST path to the goal?" Challenges unnecessary complexity. Flags technical debt.
βš”οΈ

The Adversary

glm-5.1:cloud
Minimize false positives. "Does this actually WORK? What can we CUT?" Cost-benefit, timeline realism, confidence gate enforcement.
πŸ”΅

The Synthesizer

deepseek-v3.2:cloud
Minimize information loss. "Where's the PROOF?" Preserves metadata vectors. Reviews statistical claims and validation.
πŸƒ

The Wildcard

minimax-m2.7:cloud
Minimize local minima. "What are we MISSING?" Triggers EXPLORE mode when stuck. Challenges framing, identifies blind spots.
Solved

Problems β†’ Solutions

Every problem we identified in building this system, and the solution that's already built. This is why it can be a template.

⚠️ Problem

CEO defaults to worker mode

Agent checks its own task queue instead of surveying the company. Paperclip's heartbeat prompt overrides custom instructions.

βœ… Solution

5-layer enforcement

SOUL.md + AGENTS.md + HEARTBEAT.md + cron prompt + Paperclip instructions all say "YOU ARE NOT A WORKER." All 5 must be present or it falls back.

⚠️ Problem

Hermes brain does the work

The brain has full terminal/file/web power β€” it CAN research and code. But that's a worker mindset. The brain should POINT, not BUILD.

βœ… Solution

Golden Rule encoded

"Hermes brain points, Paperclip hands execute." Pattern: identify root cause β†’ encode in agent SOUL β†’ wake the right agent β†’ verify outcome.

⚠️ Problem

Context pressure kills agents

Long conversations fill the context window. Agent loses early context, makes worse decisions.

βœ… Solution

Russian Doll Protocol

Don't compress β€” UNPACK. Child inherits PURPOSE, not history. Max depth 2. Peer-to-peer graph, not tree. Lifecycle: CREATE β†’ WORK β†’ MERGE β†’ PAUSE.

⚠️ Problem

Agents accumulate misaligned self-models

Supervisors that accumulate observations can consolidate them into misaligned self-models, even when told they're being tested (Anthropic, June 2025).

βœ… Solution

Anti-corruption inoculation

Do not audit yourself; be externally audited. The Seneschal does not score the Seneschal. The auditor is a different process. Stay Neo β€” enter any room, don't become it.

⚠️ Problem

Council groupthink

Multiple AI models from the same provider have Οβ‰ˆ0.7 error correlation. "Unanimous" isn't really independent.

βœ… Solution

Different models, different loss functions

4 different providers. Each seat optimizes a different objective (path length, false positives, information loss, local minima). Progressive dialectic: independent proposals before cross-critique.

⚠️ Problem

Shallow briefs waste agent runs

"Fix the thing" β†’ agent has no context β†’ wastes entire run producing nothing useful.

βœ… Solution

Deep Brief anatomy

Every brief includes: exact file path, specific line/function, expected output shape, acceptance criteria, model constraints. Brief depth = agent success rate.

⚠️ Problem

Duplicate issues clog backlog

Same work appears 3 times. CEO wastes every pulse re-reading duplicates.

βœ… Solution

Active dedup in survey phase

Cross-reference backlog against done/in-progress. Cancel duplicates immediately with PATCH /api/issues/{id} {status: cancelled}.

⚠️ Problem

Chi-square p > 0.05 at large N

At N > 1000, chi-square detects trivially small deviations. 0.5% difference produces p < 0.001. Causes false failures.

βœ… Solution

Cohen's w threshold

Gate on Cohen's w < 0.3, not p-value. Keep chi-square for diagnostics only. w < 0.1 = negligible, < 0.3 = acceptable, < 0.5 = investigate, β‰₯ 0.5 = reject.

⚠️ Problem

Orphaned agents running forever

Specialists created for one task never shut down. Costs accumulate.

βœ… Solution

TTL + guaranteed destroy

Every child has time-to-live. Auto-pause after 30 min inactivity. Auto-destroy after 24h idle or task complete. Lifecycle is mandatory, not optional.

The Ultimate Goal

What This Replaces

This isn't theory. Every role below is currently done by a human at a company. The ATLAS architecture replaces each with a persona that runs 24/7, costs cents per hour, never tires, and improves every iteration through compound learning.

Human RoleWhat They DoAgent ReplacementCost Ratio
CEO / General ManagerStrategy, delegation, hiring, performance reviewCEO Agent + High Council~1000:1
Product ManagerPRDs, acceptance criteria, roadmapProduct Team Agent~500:1
Software DeveloperCode features, fix bugs, review PRsBuilder Agent + QA Agent~200:1
Data AnalystFind sources, validate, build pipelinesScout + Validator Agents~300:1
QA EngineerTest suites, validation gates, regressionQA Agent (adversarial GAN)~400:1
ResearcherMarket research, competitor analysis, reportsSpecialist (unpacked on demand)~500:1
Compliance OfficerRegulatory checks, risk assessmentCompliance Specialist~300:1
Strategy ConsultantAnalysis, recommendations, frameworksHigh Council (4 seats)~1000:1
Project CoordinatorTask tracking, status reports, blockersCEO Pulse (every 30 min)~800:1

Cost ratio = human annual salary : agent compute cost per year. Based on NZ market rates vs Ollama/local + cloud API usage. Agent costs include model inference, infrastructure, and API calls. Human costs include salary, overhead, training, management, and replacement.

The Blueprint

Pipeline Specification

The complete adaptive persona pipeline β€” the JSON that started this conversation, completed and fixed. This is the spec that powers the cognitive loop at every layer.

πŸ“„ adaptive-persona-pipeline.json v2.0 Β· 21KB
Loading...

ATLAS β€” Adaptive Persona Agent Architecture
Built on Hermes Agent Β· Seneschal Pattern Β· CoALA Cognitive Architecture
Self-unpacking company Β· Russian Doll Protocol Β· High Council Governance
This is the template. Every layer is a persona. Every persona can spawn new personas. The recursion is the architecture.

← Back to Festive Hub