sigma-aerc 0.1.2

Generic AERC multi-agent collaboration framework — zero domain knowledge baked in

Σ (Sigma) — Generic AERC Multi-Agent Collaboration Framework

Analyze → Execute → Review → Converge loop with cross-review, convergence detection, and complexity-adaptive routing.

Zero domain knowledge baked in. Inject everything via SigmaConfig + ToolSpec. Works with any OpenAI-compatible LLM.

Installation

pip install -e .
# or from PyPI (future):
# pip install sigma-framework

Requires Python ≥ 3.12.

Quick Start

from sigma import (
    SigmaOrchestrator, SigmaConfig,
    AgentSpec, ToolSpec, BaseTool,
    UniversalBackend,
)

# 1. Define your domain
config = SigmaConfig(
    project_name="My Project",
    creed="We build things that work.",
    domain_keywords={
        "engineering": ["design", "calculate", "optimize"],
        "review": ["check", "verify", "audit"],
    },
    role_map={"engineer": "Lead Engineer", "reviewer": "Reviewer"},
    domain_agent_map={"engineering": "engineer", "review": "reviewer"},
)

# 2. Define your tools
class Calculator(BaseTool):
    name: str = "calculator"
    description: str = "Performs calculations"

    def _run(self, **kwargs):
        return {"success": True, "performance": {"result": 42}}

tools = {"calculator": ToolSpec(name="calculator", instance=Calculator())}

# 3. Define your agents
agents = {
    "engineer": AgentSpec(
        name="Engineer", role="Engineer",
        goal="Design solutions", backstory="Senior engineer",
        skill_files=[], tool_names=["calculator"],
        tool_instances=[Calculator()],
    ),
}

# 4. Run
backend = UniversalBackend(api_key="sk-...")
orchestrator = SigmaOrchestrator(
    config=config, agents=agents, tools=tools,
    llm_backend=backend, max_rounds=3,
)
result = orchestrator.run("Design a simple widget")

Architecture

Founder (you)
  ▼
┌─────────── AERC Loop ───────────┐
│  A (Analyze) — Agents analyze independently
│  E (Execute) — Parallel tool execution
│  R (Review) — Cross-review + devil's advocate
│  C (Converge) — Convergence judge + consensus estimation
└──────────────────────────────────┘
  ▼
  Output (REPORT.md + result.json)

Key Mechanisms

• Cross-review: Agents analyze independently (cannot see each other), then review each other's work
• Convergence judge: Numeric params Δ < 5% + no new qualitative conflicts → converged. Oscillation detection halts early.
• Complexity-adaptive: Pure rule-engine scores tasks 0-10, selects LITE/STANDARD/RIGOROUS tier
• Consensus estimation: When tools are unreliable, multiple agents estimate independently → converge to consensus range

Three Complexity Tiers

Tier	Score	Agents	Rounds	LLM Calls
LITE	≤ 2.5	2-4	1	~8
STANDARD	≤ 6.0	4-6	2-3	~30
RIGOROUS	> 6.0	8	≤ 4	~80

Configuration

All domain knowledge is injected via SigmaConfig:

Field	Purpose
project_name	Display name in reports
creed	Injected into every agent's system prompt
domain_keywords	Used by complexity assessor to detect domains
action_weights	Weights for action verbs in complexity scoring
role_map	Maps agent file stems to display names
domain_agent_map	Maps domain keys to agents for tier selection
default_tool_params	Fallback params for tools
reasonable_ranges	Expected output ranges for tool validation

Running Tests

pytest tests/ -v
pytest --cov=sigma --cov-report=term

985 tests, 80% coverage on all pure-logic modules.

Concepts

• No domain knowledge in the framework: Rocket propellants, medical terms, financial models — all injected via SigmaConfig
• Tool duck-typing: Any class with _run() is a tool — no specific base class required
• Universal LLM backend: Works with any OpenAI-compatible API (DeepSeek, GLM, Qwen, etc.)
• Immutable state: Every state update returns a new copy

License

MIT