Athanor: A Self-Sustaining Agent Workflow Orchestrator Reshaping AI Collaboration with Adversarial Planning

Athanor is a Claude Code plugin that innovatively solves traditional AI tool pain points (context bloat, single model limitations) through thin leader architecture, dual-model adversarial planning, and self-learning mechanisms. It aims to build a self-sustaining, evolving intelligent workflow system.

Named after the alchemical 'Xianzhe Furnace' (self-sustaining furnace), Athanor's core goal is to create a workflow system that grows smarter with use. Traditional AI tools face two main issues: 1) Main session context bloat leading to performance decay; 2) Plan quality limited by single model boundaries. Athanor addresses both via innovative architecture and planning mechanisms.

The thin leader (main session) never executes work directly; it only parses user input, distributes tasks to sub-agents, collects result briefs, and presents outputs. This keeps its context lean. Athanor defines 7 specialized agents:

Agent Name	Model Used	Core Responsibility
researcher	sonnet	Brainstorming research + devil's advocate perspective
analyst	sonnet	Fast parallel analysis
planner	opus	Implementation plan design
critic	opus	Plan synthesis and review
executor	sonnet	Code execution and verification loop
learner	sonnet	Session learning extraction
cleaner	haiku	Memory decay and cleanup

Athanor's innovative adversarial planning process: Input → Planner A (standard plan) → Reviewer B review; Input → Planner B (reverse plan) → Reviewer A review; Critic synthesizes both to form final plan. This 'red-blue confrontation' mechanism improves plan quality by balancing innovation and rigor, overcoming single model limitations.

Athanor's workflow has 5 stages: /athanor:discuss (decision brainstorm), /athanor:analyze (parallel analysis), /athanor:plan (adversarial planning), /athanor:work (task execution). Planning stages do not modify files; only work stage does. Execution modes:

• Serial (Solo): One task at a time, clean context (for dependent tasks).
• Parallel (Team): Group tasks into waves based on dependencies; same wave runs in parallel, with 'discovery relay' between waves.

Post /athanor:work session: Learner agent extracts structured lessons; Cleaner agent applies memory decay (time + access count). Memory uses double layers:

• Permanent layer: Architecture decisions, key patterns (never deleted).
• Work layer: Task-specific details (auto-cleaned). This allows Athanor to learn user habits and project specifics over time.

File Organization: All communication uses .athanor/sessions/{id}/ markdown files (e.g., research-a.md, plan.md, lessons/). Configuration: Customizable via athanor.json (e.g., codex.enabled, work.defaultMode, memory.decayDays). Use Cases: Complex multi-step tasks, long-term project maintenance, high-quality code review, team collaboration.

Athanor represents an important attempt to evolve AI tools into self-sustaining systems. Its core value lies in optimizing AI collaboration via architecture design (not just model power). It solves context bloat, breaks single model limits, and enables continuous evolution. It's a promising paradigm for developers seeking high-quality AI assistance.