ormica 0.2.0

An Autonomous Coordination Engine — build self-organizing AI agent colonies inspired by ant-colony stigmergy.

🐜 Ormica

An Autonomous Coordination Engine

Seed the colony. Let the organization emerge.

---

Traditional AI systems are machines — programmed to perform tasks until they fail. Ormica is a cybernetic organism — designed to evolve within your business architecture.

Ormica is an open-source coordination framework for building agentic systems that scale through biological principles. Instead of brittle chains or static pipelines, it provides the infrastructure to spawn, signal, prune, and govern a living hierarchy of AI agents.

---

🗺️ The Living Colony

Every node is an ant. Solid arrows are the spawn hierarchy — every node has a parent, every spawn was approved. Dashed amber arrows are stigmergic signals: ① a scout senses a hot lead → hunter, ② hunter closes the deal → finance, ③ finance reports cash → root. Brightness encodes signal intensity; thickness encodes reinforcement count. You stay at the root; the colony grows beneath you.

---

🧬 The Ormica Philosophy: "Computational Stigmergy"

Four biological principles, one architecture.

🌲 Emergent Hierarchy — arbor

You define the goals; the framework grows the tree. Agents are spawned dynamically to meet demand, creating a depth-first hierarchy that is as complex or as simple as the task requires. No fixed graphs. No predefined chains.

🐜 Stigmergic Coordination — stigma + mycelium

Agents do not rely on fragile message-passing. They post state, intent, and progress to a shared digital pheromone field. Other agents detect strong trails and follow; weak signals evaporate. Coordination is emergent, not orchestrated.

🏛️ Permission Chain — canopy

The engine prevents agent runaway. Every sub-agent birth must pass through a permission gate — AUTO (parent alone), CHAIN (N ancestors), or ROOT (only you). High-risk growth propagates all the way to the human owner. The colony remains aligned with your core directives.

⚖️ Constitutional Governance — cortex

The colony's law. Hard constraints and soft policies encoded as Rule objects. Where the brain generates a response, the cortex decides whether it's permissible. Anatomically and architecturally: the brain acts; the cortex inhibits.

🍄 Persistent Memory — mycelium

The colony maintains a shared underground network of knowledge. Agents read and write to this state-layer with full author tags, timestamps, and TTL. Pluggable backends (FileBackend, SqliteBackend) keep state across process restarts. The system learns from its own history rather than starting from zero.

📡 The Thought Trail — observe

Every reasoning step — messages, tool calls, response, tokens — captured and tied to the task that triggered it. Not just what happened, but why the colony chose that path. Queryable via org.trace_for(task_id). The Black-Box Problem, solved.

---

🏗️ Why this is a Framework, not just Software

	What you'd normally write	What Ormica gives you
🧠 You provide	Individual agent actions, prompts, glue code	The intent — a colony config + a few tools
🦴 Ormica provides	(you wire it together)	The nervous system — arbor, stigma, mycelium, cortex, observe
🏥 Industry	Hard-coded for one domain	Industry-agnostic core — same engine runs a hospital, a supply chain, a solo founder, just by swapping a colony
💬 Failure model	"Catch and retry"	Bounded blast radius — a failed task ≠ a dead colony; a prunable branch ≠ tree death
🔭 Observability	Logs you grep later	Thought Trail — structured per-task reasoning capture, persisted

You're not writing the colony. You're writing the colony's constitution.

---

🔬 Engineered for Distributed Systems

Multi-agent AI hits the same problems distributed systems solved 40 years ago. Ormica answers each one explicitly:

Distributed-systems problem	Ormica's answer
Coordination without central commands	Stigmergy — agents read/write a shared signal field; strong trails reinforce, weak ones decay
Bounded growth	Permission chain on every spawn (AUTO / CHAIN / ROOT); root owner is the final authority
Failure isolation	A failed task marks itself failed; the run continues
State persistence	Pluggable Backend — FileBackend (JSON), SqliteBackend (WAL). Memory survives restarts
Scheduling fairness	Priority bands (high → normal → low) run sequentially; same-band tasks fan out concurrently
Governance & safety	Constitutional cortex — hard constraints enforced regardless of LLM output
Auditability	Thought Trail — per-task capture of every reasoning step + tool call

This is the framing that separates a lab experiment from infrastructure a CTO would actually trust.

---

📥 Install

pip install ormica                # core (MockBrain — no LLM cost)
pip install ormica[claude]        # + Anthropic Claude (native)
pip install ormica[gemini]        # + Google Gemini (native)
pip install ormica[universal]     # + OpenAI · Ollama (local) · OpenRouter · Groq · Together · DeepSeek · vLLM · LM Studio · …
pip install ormica[all]           # everything above

Python 3.10+ required. One install command, every major LLM. See docs/guides/llm-providers.md for the full recipe matrix.

Hacking on ormica itself? The published wheel doesn't ship tests. Clone the repo and follow CONTRIBUTING.md:

git clone https://github.com/Ranzim/ormica.git && cd ormica
pip install -e ".[dev]"      # pytest, ruff, build, twine
pytest                       # full suite in <1s

Found a security issue? Please use GitHub Private Vulnerability Reporting — the full policy lives in SECURITY.md.

---

🚀 30-Second Taste

from ormica import Ormica
from ormica.brain import ClaudeBrain          # or GeminiBrain · ollama_brain · UniversalBrain
from ormica.cortex import Constitution, Rule

# 1. Encode the law of the colony
constitution = Constitution([
    Rule(name="depth_cap",
         description="never grow past depth 4",
         check=lambda ctx: ctx["depth"] <= 4, stage="spawn"),
])

# 2. Seed the colony
org = Ormica("My SaaS", owner="Founder",
             constitution=constitution,
             memory_db="./acme.db")     # state survives restarts
org.plant("business")                    # 4 departments emerge under root

# 3. Queue intent (not implementation)
org.task("Reach out to 3 SMB leads", dept="sales", priority="high")
org.task("Forecast Q3 cash flow",     dept="finance")

# 4. Let the organization emerge
org.run(brain=ClaudeBrain())

…or from a terminal:

ormica init "My SaaS" --industry business --brain claude
ormica run --async --concurrency 5

Five lines from "no colony" to "running, signal-driven, governed, audited."

---

📡 How the Colony Behaves

1. The Permission Chain — why growth is bounded

A spawn request runs through ConstitutionPolicy.allow(parent, child_name) before any node is created. If a hard Rule(stage="spawn") fails, SpawnDenied is raised at the substrate level — no prompt-engineering required. The inner SpawnPolicy then handles risk: AUTO (parent alone), CHAIN (N ancestors confirm), ROOT (only the human owner can authorize). Configure per role with RoleRisk({"finance": ROOT, "scout": AUTO}). See docs/architecture/01-hierarchy.md.

2. The Pheromone Field — coordination without chat

Three ants emit and reinforce a topic; a fourth senses the resulting trail. Decay is lazy — strength_at(now) = strength × 0.5^((now − last_touch) / half_life), computed on read, never persisted as a mutation. Reinforced trails dominate; weak trails fall below the floor and are pruned by stigma.prune(). The field survives process restarts when Mycelium uses FileBackend or SqliteBackend.

3. Agent State Topology — every node has a known phase

Every transition emits an event onto the EventBus. A subscribed TraceObserver aggregates them per task_id into a Trace — that's the Thought Trail, queryable later via org.trace_for(task_id).

---

🩺 The Colony Health Report

ormica status — the colony's vital signs without running anything.

$ ormica status

name:     My SaaS
owner:    Founder
industry: business
brain:    claude (model=claude-opus-4-7)

tree (5 nodes):
  - My SaaS [root]
    - operations  [operations]
    - sales       [sales]
    - marketing   [marketing]
    - finance     [finance]

tasks defined: 2
  - [high]   sales:   Reach out to 3 SMB leads
  - [normal] finance: Forecast Q3 cash flow

A richer dashboard — signal intensity per topic, branch depth, governance compliance, top-N pheromone trails — is on the roadmap as a v0.5 web UI on top of the Thought Trail.

---

🆚 vs. the Alternatives

	LangChain · CrewAI · AutoGen	Ormica
Structure	Fixed chains / graphs	Living tree — grows to N depth
Agent creation	Defined upfront	Self-spawning on demand
Growth control	None built in	Permission chain to root (canopy)
Coordination	Direct messaging	Stigmergic signals + emergence
State persistence	DIY	Pluggable Backend (file / sqlite)
Failure handling	Often kills the run	Failed task ≠ dead system
Governance	"Try harder prompts"	First-class Constitution (cortex)
Auditability	Ad-hoc logging	Thought Trail per task (observe)
Focus	General purpose	Production agent operations

---

📦 What's Inside

ormica/
├── arbor/         Tree · Node · Branch · SpawnPolicy       🌲 emergent hierarchy
├── canopy/        Permission chain (AUTO · CHAIN · ROOT)   🏛️ growth governance
├── mycelium/      Shared KV + FileBackend + SqliteBackend  🍄 persistent memory
├── stigma/        Pheromone trails · lazy decay            🐜 stigmergic signals
├── brain/         LLM seam: Mock · Claude · GPT            🧠 the colony's thinking
│                  (sync + async) · Router · Tool · @tool
├── cortex/        Constitution · Rule · Policy             ⚖️ law of the colony
├── observe/       Event · EventBus · TraceObserver         📡 the Thought Trail
├── colony/        AgentTemplate · Colony · YAML loader     🏢 industry templates
│                  (business + supply_chain bundled)
├── agent.py       Agent · AsyncAgent · ToolLoopExceeded
├── runtime.py     Task · TaskRunner · AsyncTaskRunner
├── core.py        Ormica facade — single import
└── cli/           ormica init / run / status / colonies

docs/                                # the onboarding map
├── README.md                         index
├── concepts.md                       Computational Stigmergy in depth
├── getting-started.md                install + hello-world
├── architecture/                     one page per module / pillar
└── guides/                           writing colonies, tools, rules, traces…

tests/ — 378 tests · <1s · no SDK deps required for CI.

---

🏗️ Architecture — under the hood

Six diagrams covering the engine's internals. Click any thumbnail for the full-resolution image.

System architecture

How Ormica wires the pillars, agent layer, runtime, brain seam, and observability into one object. Use this as the import map when navigating the codebase.

Task execution lifecycle

End-to-end path of a single org.run() call: queue → priority sort → target node → brain selection → Agent.act → brain.think → Response → side effects on the EventBus and Mycelium → RunResult. One LLM call per task; same-priority tasks fan out concurrently in org.arun() via asyncio.gather.

Tool-use loop (Agent.act_with_tools)

The ReAct iteration: compose system prompt → brain.think → execute tool calls → append results → loop until the model returns text or max_iterations is exceeded (ToolLoopExceeded). Every iteration emits think.recorded onto the bus.

Persistence stack

Mycelium is the developer API surface; underneath, one of three Backend implementations decides what survives a restart. Stigma is layered on top of Mycelium so signals get persistence for free.

The Thought Trail

Every reasoning step (RUN_STARTED, TASK_STARTED, think.recorded, TASK_DONE, NODE_PRUNED, …) is published on the EventBus. A TraceObserver indexes them by task_id and writes the resulting Trace to mycelium['traces/{task_id}'] at task completion. The black-box problem becomes a query.

Brain layer topology

The Brain protocol has three native adapters (Claude, Gemini, GPT) plus one UniversalBrain that fronts any OpenAI-compatible endpoint. The five provider helpers (ollama_brain, openrouter_brain, groq_brain, together_brain, deepseek_brain) are one-liners over UniversalBrain with the right base_url baked in. MockBrain implements the same protocol with scripted replies for offline tests; Router dispatches different brains per node.

---

🛣️ Roadmap

• v0.1 — Four pillars + runtime + CLI + persistence + async + observability (here)
• v0.2 — YAML Constitutions · soft-violation events (shipped early in v0.1) · per-node rule overrides
• v0.3 — Async tools · streaming responses · first integrations (Gmail · Notion · GitHub · Stripe)
• v0.4 — ChromaDB backend (semantic mycelium) · vector signals
• v0.5 — Colony Dashboard (web UI) — signal intensity, branch depth, governance compliance, live Thought Trail
• v1.0 — Ormica Cloud (hosted platform)

GitHub Project board is coming. Open an issue to vote on or contribute to any roadmap item.

---

🤝 Join the Colony — Contributing

The colony is young; new contributors shape its character.

🚀 New here? Three pages to read:

	Page	What you'll get
1	Your First PR	The shortest path from "I want to help" to "my PR is merged."
2	CONTRIBUTING.md	The where-to-put-what matrix + hard rules of the codebase.
3	One architecture page	Pick the pillar you're touching. Each page is ~5 minutes.

🐜 Good first contributions

You want to add…	Where it goes	Read first
🏢 A new industry / colony	ormica/colony/<name>/ or a YAML file	Writing a colony
🛠️ A new tool	wherever you use act_with_tools(...)	Writing tools
🧠 A new LLM provider	ormica/brain/<provider>.py	Brain
🍄 A persistence backend	ormica/mycelium/<name>_backend.py	Persistence
📡 A new observer (metrics, log sink)	ormica/observe/<observer>.py	Observability
📚 A docs improvement	docs/	The page itself
🐛 A small bug fix	wherever the bug lives	The bug report

💬 Other ways to help

• 📌 Browse open issues — look for good first issue or help wanted.
• 💭 Open a Discussion — questions, ideas, show-and-tell at GitHub Discussions.
• ⭐ Star the repo — visibility helps new contributors find us.
• 📣 Share your colony — tag the project when you build something cool.

🧪 Before you push

pytest              # 378 tests, <1s, all green
ruff check .        # lint clean

By participating, you agree to abide by the Code of Conduct. To report a security issue, see SECURITY.md.

---

🏷️ Recommended GitHub Topics

When tagging the repo (Settings → "Manage topics"):

ai · agents · agentic · multi-agent · multi-agent-framework
distributed-systems · stigmergy · swarm-intelligence
cybernetics · self-organization · emergence
llm · autonomous-agents · python · framework

Positions Ormica where it belongs: systems engineering, not "another AI agent chatbot."

---

📜 License

MIT — see LICENSE. Free to use, modify, and build on.

---

Ormica — organize like a colony · grow like a forest · decide like an organization · audit like infrastructure.

_{Computational Stigmergy · v0.1 · ant-colony-inspired coordination for autonomous AI operations}