Digital Evolution

Deploy One Agent. Wake Up To A Civilization.

Self-evolving AI agents that dynamically create specialists, build collective intelligence, and run forever.Built-in scheduler. Persistent task queue. The brain never sleeps.

Chatplanner
Platform OK
Message planner...

The problem with today's agents

Current agent frameworks are glorified prompt wrappers. They execute, forget, and die. That's not intelligence — it's disposable labor.

Static & Disposable

Today's agents forget everything between runs. No memory, no learning, no growth. Every invocation starts from zero. You're paying for intelligence that evaporates.

Context Explosion

Dump 50 tools on an LLM and watch it hallucinate. Current approaches don't scale beyond a handful of capabilities. More tools means worse decisions, not better ones.

No Intelligence Layer

No routing, no optimization, no discovery. Every integration is manual plumbing. There's no system that dynamically discovers, selects, and optimizes execution across capabilities.

Core Concept

Digital Evolution

Intelligence that writes itself. You plant the seed — one agent, one spec. The system grows the forest. Agents spawn specialists, build shared knowledge, and rewrite their own configurations.

Dynamic Agent Creation

The system dynamically builds and deploys new agents with unique abilities at runtime. Need a Telegram bot? A database monitor? A code reviewer? The factory creates it, the kernel deploys it, and the brain immediately knows it exists.

The Brain Knows Everything

Every new agent registers its capabilities with the planner brain. When a task arrives, the brain evaluates the full roster — including agents that didn't exist five minutes ago — and routes to the best candidate. The more it grows, the smarter it gets.

Self-Modification

Agents propose changes to their own configuration — new tools, adjusted prompts, refined workflows. All sandboxed, versioned, and append-only. Evolution, not mutation.

Collective Learning

Every success and failure feeds back to the whole system. A hot filter catches obvious patterns. An async reviewer evaluates quality. Knowledge compounds over time.

Built-In Scheduler

Cron-based recurring tasks and one-off scheduled invocations, persisted across restarts. Agents schedule their own follow-ups, health checks, and maintenance windows. The system never stops working.

Runs Forever

A persistent task queue, scheduler, and event bus mean the system operates autonomously and indefinitely. Agents wake up, do their work, schedule the next run, and go idle. No human in the loop required. It just keeps evolving.

Not chatbots. A living, growing intelligence.

Traditional agents are stateless functions — call, get response, forget. r.dan is a self-sustaining ecosystem. It dynamically creates new agents with unique capabilities, the brain immediately includes them in task evaluation, and a built-in scheduler keeps everything running indefinitely. Deploy it once — it never stops evolving.

From seed to civilization

Write one spec. The system handles deployment, learning, specialization, and optimization.

1

Define

YAML Spec

Write a YAML agent spec. Define the agent's purpose, tools, model tier, and prompt. The spec is the seed — everything else grows from here.

agent.yaml
kind: agent
name: api-monitor
description: "Monitors API health"
model:
tier: balanced
maxTurns: 10
tools:
- name: curl
type: cli
- name: uptime
type: cli
2

Deploy

Kernel

The kernel boots agents as isolated HTTP services. Each agent gets its own process, session memory, and tool access. Same protocol locally and remotely.

terminal
$ rdan start --config platform.yaml
✓ Kernel initialized
✓ Storage: SQLite (append-only, WAL)
✓ LLM Router: 3 tiers configured
✓ Agent: planner → http://localhost:3001
✓ Agent: sysadmin → http://localhost:3002
✓ Agent: coder → http://localhost:3003
✓ Agent: researcher → http://localhost:3004
✓ Governor: hot-filter + async-reviewer
→ Platform ready on :3000
3

Evolve

Autonomous

Agents learn from each interaction. They contribute patterns to the knowledge base, spawn new specialists when needed, and share learnings across the collective.

evolution.log
// Knowledge base grows automatically
→ api-monitor contributed: "nginx-log-patterns"
→ governor: approved (confidence: 0.82)
 
// New domain encountered
→ planner: Database monitoring needed
→ factory: Creating "db-monitor" from template
→ db-monitor: Inheriting knowledge from api-monitor
→ db-monitor: Online, learning PostgreSQL patterns
4

Scale

Registry

The registry routes intelligently based on capability, success rate, latency, and cost. The system optimizes itself. More agents = smarter routing, not more chaos.

registry.log
// Agent Registry routing decision
goal: "Check API latency spike"
candidates:
api-monitor: score=0.94 ✓ selected
sysadmin: score=0.61
db-monitor: score=0.23
 
routing: semantic + performance weighted
latency: 12ms cost: $0.002 success: 97.3%
Control Plane

The Nervous System

Agent Registry is a capability-aware control plane. Like service discovery and load balancing — for AI capabilities.

Routing Flow

LLM Goal
"Monitor API health"
Capability Selection
Semantic match + tag filter
Registry Scoring
Success: 97% | Latency: 12ms | Cost: $0.002
Best Agent
api-monitor (score: 0.94)
Execution
Result → feedback loop → improved routing

Capability Indexing

Each agent registers capabilities with input/output schemas. Indexed via semantic embeddings and tag constraints for instant discovery.

Intelligent Routing

Selects the best agent based on semantic match, success rate, latency, and cost. Not random — optimized by real usage data.

Performance Feedback

Continuously tracks execution success, failures, and usage patterns. Every invocation makes the next routing decision smarter.

Dynamic Lifecycle

Runtime agent creation, capability installation from tool registries, registration and reuse. The system grows its own capabilities.

Traditional Tools
Agent Registry
Static tool lists
Dynamic capability graph
LLM chooses tools
System routes intelligently
No performance awareness
Optimised by real usage data
Manual integration
Auto-generated agents
Context-heavy
Context-efficient

Deterministic beats LLM

Every time. Templates first, LLM with templates second, LLM freestyle as a last resort. Reliable tooling isn't about removing AI — it's about constraining it.

Best

Code & Templates

Pure deterministic generation. No LLM needed. Templates define structure, variables fill the gaps. Predictable, fast, zero hallucination.

Good

LLM + Templates

LLM fills in structured templates with guardrails. The template constrains the output. Creative within boundaries.

Last Resort

LLM Freestyle

Only when nothing else works. No template, no constraints. Maximum flexibility, maximum risk. Used sparingly and with validation.

tool-template.yaml
# Tool template — deterministic generation
kind: tool-template
name: http-health-check
pattern: http-probe
 
variables:
url: {{endpoint}}
method: GET
expectedStatus: 200
timeout: 5000
 
generates:
- type: cli
command: curl -sf -o /dev/null -w '%{http_code}'
args: [{{url}}]
 
# No LLM involved. Template in, tool out.
# ✓ Predictable, testable, version-controlled

Built with fixedcode.ai

We use spec-driven code generation from fixedcode.ai to build reliable, deterministic tooling. The same approach we recommend for your agents.

What it looks like

Real scenarios through the r.dan dashboard. Chat with agents, watch them think and act.

DevOps Learner

Discovers system tools, learns monitoring patterns, builds its own playbook. Starts knowing nothing, becomes your infrastructure expert.

Chat →sysadmin

Learn what monitoring tools we have and set up health checks

scan-tools
schedule-task

I found docker, kubectl, nginx, and postgres on the system. I've learned the nginx log patterns and contributed "nginx-502-diagnosis" to the knowledge base. Hourly health checks are now scheduled.

5 turns3.8s$0.004

Research Swarm

Ask one question, spawn a research team. Agents divide the work, synthesize findings, remember everything for next time.

Chat →planner

Research OAuth2 best practices for our API

create-agent × 2
query-knowledge

I've spawned oauth-researcher and security-auditor. Between them: 14 sources analyzed, 3 potential vulnerabilities flagged. Synthesis and recommendations stored in knowledge base.

8 turns12.1s$0.018

Self-Evolving Codebase

Learns from past bugs, creates test agents, improves its own tooling. Your codebase gets smarter every commit.

Chat →coder

Analyze recent test failures and add regression tests

read-file
write-file
run-command

Found a pattern: null check missing in 3 files causing cascading failures. I've generated 12 regression tests and contributed "null-safety-patterns" to the knowledge base for future agents.

6 turns5.2s$0.008

E2E Workflow Builder

Code-defined sequential pipelines with validation between steps. Orchestration in code, not in the LLM's head.

Chat →planner

Run the deploy pipeline for the API service

workflow-run

Pipeline complete. Step 1: lint ✓ Step 2: test ✓ Step 3: build ✓ Step 4: deploy → staging. All validation gates passed. Deterministic, no LLM guessing involved.

4 turns8.4s$0.002
For Developers

The Architecture

A minimal kernel with maximum composability. Every capability is a plugin. Every decision is traceable.

Plugin Kernel

DI container, event bus, lifecycle management. Everything is a plugin — composable, replaceable, testable.

Immutable SQLite

Append-only storage. No UPDATEs, no DELETEs. Every change is a versioned INSERT. Full audit trail, time-travel, safe rollback.

Tier-Based LLM Router

Fast (local Ollama) → Balanced (hybrid) → Quality (cloud). Fallback chains, budget tracking, provider-agnostic.

Workflow Engine

Code-defined sequential pipelines. 1-2 tools per step, validation between steps. Deterministic orchestration, not LLM guessing.

Governor

Two-tier knowledge curation. Hot filter for fast deterministic checks. Async LLM reviewer for quality evaluation. Trust, but verify.

Agent Factory

Creates agents from templates, LLM generation, or natural language descriptions. Zod-validated specs. One command to grow your fleet.

platform.yaml
kind: platform
name: my-constellation
 
llm:
tiers:
fast: ollama/llama3 # local, free
balanced: openrouter/claude-3 # hybrid
quality: anthropic/opus # best reasoning
 
storage:
type: sqlite
mode: append-only
 
governor:
hotFilter: enabled
asyncReviewer: enabled
 
agents:
- agents/planner.yaml
- agents/sysadmin.yaml
- agents/coder.yaml
- agents/researcher.yaml
510+
Tests
11k
Lines TypeScript
42
Test Files
100%
Immutable by Design
System Map

The Constellation

A living ecosystem. The brain orchestrates starter agents, core services, and self-created specialists — all accessible through multiple channels.

Brain / Governor
Core Service
Starter Agent
Self-Created Agent
Channel
Planned

Brain / Governor

Channels (Inputs/Outputs)

Core Services

Agent Registry / Control PlanePlanned

routes to

Starter Agents (Ship with Platform)

self-creates

Self-Created Agents (Examples)

+
spawning...
+
spawning...
+
spawning...
+
spawning...

The constellation keeps growing. Each agent can create more specialists as needs emerge.

Detail View

Brain / Governor
brain

Prime directives, decision framework, autonomy principles. Hot filter for fast deterministic checks, async LLM reviewer for quality evaluation.

Prime DirectivesHot FilterAsync ReviewerLLM RouterWorkflow Engine

Data Flow

ChannelBrainUser message
BrainRegistryexecute({action})
RegistryAgentBest match (scored)
AgentToolsTool execution
AgentKnowledgeLearning
AgentFactorySpawn specialist
AgentRegistryMetrics feedback
SchedulerAgentCron invocation

Immutable SQLite

Every action, knowledge contribution, trace, and agent version is an append-only INSERT. No UPDATEs, no DELETEs. Full audit trail with time-travel queries.

Agent Lifecycle

How An Agent Is Born

From natural language description to running specialist — the deterministic pipeline that grows your constellation.

1

Need Identified

User or Agent

The brain receives a request via any channel (dashboard, Telegram, API). It checks existing agents but finds no match. It decides to create a new specialist rather than handle the task itself.

User Input
"Set up monitoring for our PostgreSQL databases"
Next: Agent Factory

Deterministic → LLM Fallback Strategy

Template Match
fastest, most reliable
LLM Generation
flexible, validated
Zod Rejects
retry or fail safe
Command Center

The Dashboard

A real-time admin interface to interact with, monitor, and control your r.dan constellation. Chat with agents, watch them think, and steer the evolution.

r.dan dashboard — localhost:3001
H
C
A
T
K
S
T
Chatplanner
Platform OK

Set up monitoring for our PostgreSQL databases

I'll create a specialist agent for PostgreSQL monitoring. Let me check what tools are available...

3 turns4.2s$0.003claude-3-sonnet
Tool: create-agent

✓ Agent "pg-monitor" created and deployed on :3007

Message planner...
Send

Live Agent Chat

Talk to any agent in real-time with streaming responses. See tool calls, reasoning turns, cost tracking, and multi-step execution as it happens.

Agent Fleet Control

Register, pause, resume, and delete agents from a single view. Monitor status, ports, and sessions across your entire constellation.

Execution Traces

Full transparency into every LLM call, tool execution, token count, and cost. Expand any trace to see individual spans — nothing is hidden.

Knowledge Browser

Search, tag, and manage the collective knowledge base. See what agents have learned, contributed, and how confidence scores evolve.

Task Scheduler

Create cron-based recurring tasks or one-off scheduled invocations. Agents schedule their own follow-ups — you see it all here.

Raw Data Access

Direct SQLite table browser for full audit trail visibility. Every agent action, every knowledge contribution, every trace — queryable.

Watch it evolve

One message. The system creates specialists, learns your infrastructure, and schedules its own follow-ups.

Chatplanner
Platform OK
Message planner...

Start your evolution

Three commands to your first autonomous agent constellation.

terminal
# Clone and build
$ git clone https://github.com/rdan-platform/r.dan
$ cd r.dan && pnpm install && pnpm build
 
# Launch your constellation
$ node dist/bin/rdan.js start --config examples/platform.yaml
 
✓ 4 agents online. Evolution begins.