Autonomous Agents (Unified Reasoning + Stem Cell)
Rnix supports autonomous agent reasoning through the unified reasoning loop and stem cell differentiation — agents that autonomously decide their actions each step, and automatically specialize based on task requirements.
Unified Reasoning Loop
Every Rnix agent runs a single reasonStep loop. Each step, the LLM autonomously selects one of seven action types:
┌─────────────────────────────────────────┐
│ reasonStep Loop │
│ │
│ LLM ──→ ActionType ──→ Execute │
│ ▲ │ │
│ └────── context ◄────────┘ │
│ │
└─────────────────────────────────────────┘| ActionType | Description |
|---|---|
tool_call | Execute a VFS tool call directly |
plan | Output an execution plan, written to context as RoleAssistant |
spawn | Create a child process (mission command) |
complete | Output final result and exit (code=0) |
specialize | Dynamically load a Skill (stem cell progressive specialization) |
replan | Revise the current plan |
text | Plain text output (final answer) |
Planning Configuration
Planning is a configurable capability, not a separate mode:
name: autonomous-analyst
description: "Self-directed code analysis agent"
planning: true # default: true — LLM can choose to plan
models:
preferred: deepseek-v4-flash
skills:
- code-analysis
- security-scanplanning: true(default) — plan/replan guidance injected into prompt, LLM can choose to plan before executingplanning: false— no plan guidance in prompt, LLM executes tools directly
Built-in Safety Mechanisms
- VFS flags auto-downgrade: empty payload →
O_RDONLY, non-empty →O_RDWR - Error injection: tool errors are injected as tool messages into LLM context, so the LLM can perceive and adjust strategy
- Circuit breaker: 3 consecutive tool_call/spawn failures triggers automatic process termination (exit code=1)
- Recoverable errors: specialize/plan/replan failures do not count toward circuit breaker (recoverable logic errors)
Mission Command
Agents can autonomously spawn child agents using mission command — specifying intent without dictating execution details:
Parent → Spawn "Analyze authentication module"
→ Child decides HOW to analyze
→ Child may spawn its own children
→ Results flow back to parent contextChild agents run their own reasonStep loops, making independent decisions about execution strategy.
Stem Cell Differentiation
Rnix provides a universal base agent (Stem Agent) that automatically specializes based on the task it receives.
How It Works
Generic Stem Agent (no Skills)
│
│ Receives intent: "Analyze security vulnerabilities"
▼
Auto-match Skills
│ → code-analysis (relevance: 0.9)
│ → security-scan (relevance: 0.95)
▼
Differentiated Agent
Skills: [security-scan, code-analysis]
Capabilities: /dev/fs, /dev/shell- Base agent starts with only core reasoning capability and unified reasoning loop, no bound Skills
- Auto-matching analyzes the intent against all available Skill metadata to find the best matches. Matched skills are then re-ranked using reputation scores and synergy matrix data — high-performing skill combinations are prioritized, with ε-exploration to prevent "rich-get-richer" lock-in
- Progressive specialization loads core Skills first, then dynamically loads additional Skills via
specializeaction as needed during execution - Differentiation Memory (DiffMemory) records intent→skills mappings and persists them as JSON Lines files (
$PROJECT/.rnix/diffmemory/). On daemon restart, historical mappings are reloaded. When a similar intent arrives, previously successful skill combinations are recalled instantly, bypassing the matching computation
Differentiation Lineage
The lineage system provides observability telemetry for the differentiation process — it records what happened, but does not influence future spawn or selection decisions. View the complete differentiation path:
$ rnix lineage <pid>
PID 5: autonomous-analyst
Base: stem-agent
Differentiation path:
1. [auto] security-scan (intent match: 0.95)
2. [auto] code-analysis (intent match: 0.90)
3. [runtime] dependency-check (loaded at step 4, tool need)Feedback Loop: Reputation → Stem Match
Differentiation is not a one-shot process — it improves over time through a closed feedback loop:
Intent → StemMatcher (keyword match)
→ Reputation + Synergy re-rank
→ Spawn with selected Skills
→ Execution result
→ RecordResult (reputation update)
→ RecordCombo (synergy matrix update)
→ Next match benefits from accumulated dataThis means the system exhibits natural selection: skill combinations that produce better results are gradually preferred, while the ε-exploration parameter ensures novel combinations are still tried.
Intent-Driven Decomposition
The Intent System (see Intent System) works naturally with autonomous agents — the Reconciler decomposes high-level intents into sub-task DAGs, and each node can be executed by an agent that autonomously decides its approach through the unified reasoning loop.
User Intent: "Refactor authentication to use JWT"
│
├── [Reconciler decomposes]
│
├── analyze → autonomously decides what to examine
├── design → plans architecture approach, spawns sub-agents if needed
├── implement → executes tools, spawns sub-agents as needed
└── test → decides test strategy, runs verificationRelated Documentation
- Intelligence Emergence — How differentiation, reputation, and synergy produce emergent intelligence
- Intent System — Declarative intent decomposition and reconciliation
- Token Economy — Budget pools, reputation, and synergy feedback
- Security & Self-Healing — Immune system and neuroplasticity
- Agents & Skills — Agent and Skill definitions
- Compose Orchestration — Static multi-agent DAG workflows