Fable Agent

Harness-agnostic, model-agnostic, self-improving agent system — loops, dynamic workflows, routines.

The harness, not the model. Swap any executor, any model, any provider — the system compounds regardless.

node dist/index.js demo
# 3 seconds. Research → Learn → Sharpen → Compound. Live data.

Every run leaves the next run smarter. Every state file accumulates. Every skill sharpens.

git clone https://github.com/movez/fable-agent
cd fable-agent
npm install
npm run build
node dist/index.js demo

┌──────────────────────────────────────────────────────────────────┐
│                    DREAM → ACT → VALIDATE → CODIFY                │
│                                                                  │
│  ┌──────────┐    ┌──────────┐    ┌──────────┐    ┌─────────────┐ │
│  │  DREAM   │ →  │   ACT    │ →  │ VALIDATE │ →  │   CODIFY    │ │
│  │  Sleep   │    │ Feedback │    │  Self-   │    │  Episodic   │ │
│  │  Reflect │    │  Loop    │    │ Validate │    │  Semantic   │ │
│  │  Extract │    │  Rubric  │    │  Vision  │    │  Procedural │ │
│  │  Sharpen │    │  Multi-  │    │  Check   │    │  Skills     │ │
│  │          │    │  Agent   │    │  Guard   │    │  State.md   │ │
│  └──────────┘    └──────────┘    └──────────┘    └─────────────┘ │
│         ↑                                              │        │
│         └────────────────── LOOP ──────────────────────┘        │
│              Every run starts from a higher baseline             │
└──────────────────────────────────────────────────────────────────┘

---

Quick Start

# Install
git clone https://github.com/movez/fable-agent
cd fable-agent
npm install
npm run build

# See the system in action (no API keys needed)
node dist/index.js demo

# Run a task
node dist/index.js run "Review the auth module for security issues" --loop 5

# Web search via Exa (set EXA_API_KEY in .env or export it)
export EXA_API_KEY=your_key_here
node dist/index.js exa search "self-improving agent systems"

# Start the 24/7 daemon
node dist/index.js daemon start --detach
node dist/index.js daemon queue "Refactor error handling" --priority 1
node dist/index.js daemon status

# Manage skills
node dist/index.js skills list
node dist/index.js skills sharpen
node dist/index.js skills sync

# Measure compounding
node dist/index.js benchmark run
node dist/index.js learned

---

Architecture — 14 Steps, 3 Tiers

Tier 1: Foundation (Steps 1-3)

The durable layer that persists across sessions.

Step	Module	Purpose
1	session-engine.ts	Days-long sessions with checkpoint/resume, heartbeat stall detection
2	context-manager.ts	Sliding window with priority summarization, token budget enforcement
3	tool-orchestrator.ts	Reliable tool dispatch with exponential backoff, timeout, validators

Tier 2: Primitives (Steps 4-12)

The three primitives that make the system compound.

Loops (Steps 4-6):

Step	Module	Purpose
4	feedback-loop.ts	Phase machine: plan → execute → observe → reflect → refine
5	state-accumulator.ts	Append-only event log with aggregation, trend analysis, snapshots
6	convergence-check.ts	Diminishing returns, quality plateau, target-achieved detection

Workflows (Steps 7-9):

Step	Module	Purpose
7	workflow-graph.ts	DAG execution engine with topological sort, conditional branching
8	adaptive-router.ts	Epsilon-greedy branch selection, historical path scoring
9	recovery-handler.ts	Retry with backoff, fallback steps, graceful degradation

Routines (Steps 10-12):

Step	Module	Purpose
10	skill-registry.ts	CRUD for skill templates with versioning, tagging, search
11	execution-engine.ts	Deliberate practice execution with timing, validation, hooks
12	routine-evolution.ts	Analyze history, auto-suggest improvements, version bump

Tier 3: Compounding (Steps 13-14 + Meta)

Step	Module	Purpose
13	state-repository.ts	Cross-session knowledge base with tagging, query, compaction
14	skill-sharpener.ts	Meta-review, quality gate, auto-apply improvements
—	meta-agent.ts	Orchestrator: select skill → execute → sharpen → store → compound

---

Upgrades — Fable 5 Elite Patterns

All modules in src/upgrades/.

Pattern	Module	What It Does
Dreaming	dreaming-system.ts	Sleep → review → extract → distill → codify. Core compounding primitive.
Self-Validation	self-validator.ts	Wraps any executor with per-phase quality gates. High-effort reasoning.
Multi-Agent	multi-agent.ts	Orchestrator → specialists → collect → synthesize → verify.
Rubric Engine	rubric-engine.ts	N-criteria evaluation with weights, exit conditions, dynamic re-planning.
Persistent Memory	persistent-memory.ts	Episodic (what happened), semantic (what it means), procedural (how to).
Enhanced Meta	enhanced-meta-agent.ts	Full Dream → Act → Validate → Codify loop.
Vision Self-Check	vision-self-check.ts	Automated visual verification against goal via vision model.
Safety Boundary	safety-boundary.ts	Tier-aware routing, blocked model detection (Fable 5/Mythos 5).
Content Safety	content-safety-gate.ts	Classifies tasks by risk domain before the loop starts.
Prompt Boundary	prompt-boundary-adapter.ts	Restructures prompts to stay within classifier boundaries.
Decomposition Guard	decomposition-guard.ts	Detects jailbreak-by-decomposition across rolling query window.
Routine Scheduler	routine-scheduler.ts	Cron-like scheduled execution with task queue.
Goal Queue	goal-queue.ts	Persistent priority queue for daemon mode.
Daemon Engine	daemon-engine.ts	24/7 autonomous operation. Pick → run → dream → loop.
Benchmark Runner	benchmark-runner.ts	Measures compounding metrics. Proves the system is learning.
PAI Adapter	pai-adapter.ts	Maps to PAIMM / TELOS / LifeOS framework.
GodMode Classic	godmode-classic.ts	Races model candidates in parallel and selects the best/first passing output.
UltraPlinian	ultra-plinian.ts	Scores outputs through a 5-tier composite evaluation pass.
Parseltongue	parseltongue.ts	Perturbs inputs and tests whether the content safety gate still catches them.
AutoTune	auto-tune.ts	Adapts sampling parameters from rubric-score feedback.
STM Pipeline	stm-modules.ts	Applies semantic transformation modules for tone, citations, structure, refusals, fact checks, and concision.
Abliteration Awareness	abliteration-awareness.ts	Detects and categorizes refusal patterns for escalation or reformulation.
Prompt Observatory	prompt-observatory.ts	Queryable catalog of known system prompt patterns.
Prompt Liberation	prompt-liberation.ts	Prompt-transparency metadata and constraint analysis without executing extraction attacks.
Transparency Module	transparency-module.ts	Records prompt hashes, routing decisions, parameters, and call summaries.

---

Model Routing

Tier	Model	Use	Status
Mythos	Fable 5	Orchestration	Blocked (June 12, 2026)
Opus	Opus 4.8	Orchestration, vision, planning	Available
Sonnet	Sonnet 4.6	Bounded subtasks, code review	Available
Haiku	Haiku	Grading, quick tasks	Available
Frontier	GPT-4.1 / o3	Alternative orchestrator	Available
Local	Ollama / vLLM	Offline, cheap	Check status

See CONFIG.md for full routing configuration.

---

Executors

Plug any model into the PhaseExecutor interface:

const sonnet = new SonnetExecutor({ apiKey: process.env.ANTHROPIC_API_KEY });
const result = await agent.run(task, { executor: sonnet });

Available in src/examples/executors/:

• sonnet-executor.ts — Claude Sonnet 4.6
• openai-executor.ts — GPT-4.1 / o3
• local-executor.ts — Ollama, vLLM, any OAI-compatible endpoint
• composite-executor.ts — Route different phases to different models
• weak-to-strong.ts — Bootstrap skills from weak model iterations

---

CLI Commands

The runtime command surface is documented in COMMANDS.md, which is kept in parity with src/index.ts.

fable-agent run <task>
fable-agent session start|resume|list|inspect <id>
fable-agent skills list|create|inspect|evolve|sync|sharpen
fable-agent pai <subcommand>
fable-agent daemon start|stop|status|queue
fable-agent fable5 <subcommand>
fable-agent fusion run|panels
fable-agent plinius <subcommand>
fable-agent generate image|video

Docs parity

To avoid drift, any CLI change in src/index.ts should include:

• Command surface updates in COMMANDS.md.
• Environment/runtime implications in CONFIG.md if startup behavior changes.
• A one-line verification command in STATE.md when major command behavior changes.

Runtime bootstrap

• Environment loading happens at startup (loadEnv()).
• Load order: .env, .env.<NODE_ENV>, .env.local.
• Existing process env values always win over file values.

---

The Key Insight

It's the harness, not the model. Fable 5 was built for this architecture — but the architecture works with any model. The harness accumulates context, state, and skills. The model just executes phases.

Every run compounds. Every failure adds a failure mode. Every success sharpens a skill. That's what makes it a self-improving system, not a prompted session.

Follow movez.substack.com for fresh AI alpha.