---

The problem with AI agents today

The math is brutal:

An agent with 85% accuracy fails 80% of the time on a 10-step task.
85% × 85% × 85% × ... × 85% (×10) = 19.7% success rate.

This isn't a bug. It's compound failure — and it's killing agent projects everywhere.

The numbers, from 2025 research:

Source	Finding
Carnegie Mellon	Best agent (Gemini 2.5 Pro) fails real-world tasks 70% of the time
Carnegie Mellon	GPT-4o fails 91.4% of office tasks
Gartner	40% of agentic AI projects will be cancelled by 2027
MIT	95% of AI pilot programs stall — zero measurable P&L impact
S&P Global	42% of companies abandoned most AI initiatives in 2024
Enterprise survey	Only 95 of 1,837 AI teams have agents actually live in production

The root cause (MIT calls it "the learning gap"):

Day 1:   Agent fails on nonce error
Day 100: Agent fails on the same nonce error
         Nothing was learned. Every failure is a fresh start.

"We're deploying goldfish and expecting them to become sharks."

VialOS fixes this:

Failure            →   PCEC engine diagnoses it
                   →   Constructs repair strategies
                   →   Picks the best one (Q-learning)
                   →   Executes and records in Gene Map
                   →   Next identical failure: instant fix
                   →   Confidence compounds over time

Quick Start

npm install @vial-agent/runtime

import { ClaudeCodeAdapter } from '@vial-agent/runtime'

const adapter = new ClaudeCodeAdapter({
  geneMapPath: './my-agent.db'  // persistent memory lives here
})

const result = await adapter.run(
  { prompt: 'Your agent task here' },
  async (prompt, ctx) => {
    // your existing agent code — unchanged
    const response = await yourAgent.run(prompt)
    return { output: response.text, success: !response.error }
  }
)

console.log(result.vialStrategy)  // what repair strategy was used
console.log(result.vialAttempts)  // how many attempts it took

That's it. Your agent now self-heals and learns from every failure.

---

How it works

VialOS wraps your agent with a PCEC loop and a Gene Map:

                    ┌─────────────────────────────────┐
                    │         Your Agent               │
                    └──────────────┬──────────────────┘
                                   │
                    ┌──────────────▼──────────────────┐
                    │         VialOS Runtime           │
                    │                                  │
                    │  P  Perceive  ← diagnose failure │
                    │  C  Construct ← generate fixes   │
                    │  E  Evaluate  ← score strategies │
                    │  C  Commit    ← execute + learn  │
                    │                                  │
                    │  ┌──────────────────────────┐   │
                    │  │  Gene Map (SQLite)        │   │
                    │  │  · Failure patterns       │   │
                    │  │  · Repair strategies      │   │
                    │  │  · Success rates (Q-value)│   │
                    │  └──────────────────────────┘   │
                    └─────────────────────────────────┘

Gene Map: persistent memory that compounds

Every execution adds to the Gene Map. Every failure teaches a repair strategy. Every success increases the Q-value of that strategy.

Attempt 1:  nonce_error → tries 3 strategies → succeeds on attempt 2  (confidence: 70%)
Attempt 2:  nonce_error → Gene Map: use increment_nonce               (confidence: 80%)
Attempt 10: nonce_error → instant fix, first attempt                  (confidence: 95%)

The longer you run, the smarter it gets. Unlike LLM context, Gene Map never forgets.

---

Examples

Wrap any Claude Code agent

import Anthropic from '@anthropic-ai/sdk'
import { ClaudeCodeAdapter } from '@vial-agent/runtime'

const client = new Anthropic()
const adapter = new ClaudeCodeAdapter({ geneMapPath: './agent.db' })

// your agent — completely unchanged
async function myAgent(prompt: string) {
  const response = await client.messages.create({
    model: 'claude-opus-4-5',
    max_tokens: 1000,
    messages: [{ role: 'user', content: prompt }]
  })
  return response.content[0].type === 'text' ? response.content[0].text : ''
}

// wrap with VialOS — one line
const result = await adapter.run(
  { prompt: 'Fix this failing transaction', maxTurns: 10 },
  async (prompt, ctx) => {
    const output = await myAgent(prompt)
    return { output, success: true }
  }
)

Payment agents (Helix integration)

import { ClaudeCodeAdapter } from '@vial-agent/runtime'

const adapter = new ClaudeCodeAdapter({
  geneMapPath: './helix-genes.db',
  maxRetries: 3,
})

// Automatically learns: nonce errors → increment_nonce
//                       gas errors  → increase_gas_limit
//                       balance err → check_balance first
const result = await adapter.run(
  { prompt: `Repair this transaction: ${JSON.stringify(failedTx)}` },
  async (prompt, ctx) => {
    const repair = await helixClient.heal({ transaction: failedTx, error })
    return { output: JSON.stringify(repair), success: repair.success }
  }
)

Check what your agent has learned

const genes = adapter.getGeneMap().list()

genes.forEach(gene => {
  console.log(`${gene.failureCode}: ${gene.strategy} (${Math.round(gene.qValue * 100)}% success)`)
})

// nonce_error:   increment_nonce    (95% success)
// gas_error:     increase_gas_limit (88% success)
// timeout:       retry_with_backoff (72% success)

Multi-agent meeting room (AMP Protocol)

import { AMPBus } from '@vial-agent/runtime'

const bus = new AMPBus()
const session = bus.createSession('Fix production bug', ['code', 'review'], 100_000)

// CodeAgent proposes a fix
await bus.publish(session.id, 'code', 'review', 'proposal',
  'I will fix the nonce error in transaction.ts', { confidence: 0.9 })

// ReviewAgent responds
await bus.publish(session.id, 'review', 'code', 'support',
  'Approved — increment nonce by current chain value', { confidence: 0.85 })

// Get all messages
const messages = bus.getMessages(session.id)

---

What's included

Module	Description
ClaudeCodeAdapter	Wrap any agent with PCEC self-healing
GeneMap	SQLite persistent memory with Q-learning
PCEC	Perceive → Construct → Evaluate → Commit engine
GeneDream	Offline knowledge consolidation (4 phases)
AMPBus	Agent Meeting Protocol — structured multi-agent communication

---

Compared to alternatives

	LangChain	AutoGen	CrewAI	VialOS
Self-healing	❌	❌	❌	✅ PCEC loop
Persistent memory	❌	❌	❌	✅ Gene Map
Learns from failures	❌	❌	❌	✅ Q-learning
Multi-agent protocol	partial	✅	✅	✅ AMP
Model agnostic	✅	✅	✅	✅
One-line integration	❌	❌	❌	✅

The difference: other frameworks coordinate agents. VialOS makes them durable.

---

Architecture

VialOS Runtime is the foundation layer. Verticals run on top:

VialOS Runtime  (@vial-agent/runtime)
      │
      ├── Helix       — self-healing payment agents   (npm: @helix-agent/core)
      ├── Your App    — bring your own vertical
      └── ...

Inspired by analysis of Claude Code internals (KAIROS, COORDINATOR, DreamTask) and independently validated by EvoMap's convergence on the same architecture.

---

API

ClaudeCodeAdapter

new ClaudeCodeAdapter(config?: {
  geneMapPath?: string      // default: './vial-claude-code.db'
  maxRetries?: number       // default: 3
  dreamThreshold?: number   // run Gene Dream after N capsules (default: 50)
  verbose?: boolean         // log PCEC steps (default: false)
})

adapter.run(config, executor) → Promise<AdapterResult>
adapter.getGeneMap()         → GeneMap
adapter.dream()              → Promise<void>  // force Gene Dream
adapter.close()              → void

GeneMap

new GeneMap(dbPath?: string)

geneMap.query(context)                → Gene[]
geneMap.getBestStrategy(failureCode)  → Gene | null
geneMap.recordCapsule(capsule)        → string
geneMap.list()                        → Gene[]
geneMap.shouldDream(threshold)        → boolean
geneMap.close()                       → void

PCEC

new PCEC({ geneMap, maxRetries? })

pcec.repair(failure, executor) → Promise<PCECResult>
// PCECResult: { success, attempts, finalStrategy, escalated }

---

Installation

# npm
npm install @vial-agent/runtime

# pnpm
pnpm add @vial-agent/runtime

# yarn
yarn add @vial-agent/runtime

Requires: Node.js 18+, better-sqlite3

---

Philosophy

"The model is the most interchangeable part. The harness is where years of production experience live." — Inside Claude Code

VialOS is the harness. PCEC + Gene Map is the experience that compounds.

---

Built with VialOS

• Helix — self-healing payment SDK for AI agents. Listed on MPPScan.

---

Contributing

PRs welcome. See CONTRIBUTING.md.

Areas we need help:

• More Perceive classifiers for different domains (LLM errors, database errors, API errors)
• More Construct strategies
• Adapters for other agent frameworks (LangChain, AutoGen, LlamaIndex)

---

License

MIT

---