🗺️ Arachne Implementation Roadmap

April 26, 2026 · View on GitHub

Status: Beta (v0.1.0) Goal: DSPy-native self-healing AI agent runtime.

👁️ Vision

Arachne is an autonomous agent runtime designed to replace brittle prompt chains with dynamic, DSPy-native agent graphs. It is built for long-running reliability, featuring autonomous self-correction, crash recovery, and strict protocol-first tool governance using the Model Context Protocol (MCP).

🏗️ Core Philosophy

  • Thin Orchestration, Thick Intelligence: Minimal structural scaffolding; all "intelligence" lives in DSPy-native modules (Weaver, Evaluator, Healer).
  • Self-Healing Initial Action: The system automatically detects failures, diagnoses causes, and re-weaves the graph before human intervention.
  • Protocol-First Ecosystem: Seamlessly connect to any MCP server to safely expand agent capabilities.
  • Stateful Persistence: Every wave and node result is checkpointed to disk for recovery and observability.

🏆 Completed Milestones

✅ Phase 1: Foundations & Tooling

  • MCP Client: Dynamic conversion of MCP tools to dspy.Tool.
  • Pointer Pattern: Spillover protection for massive tool outputs (>30KB).
  • Security: Strict command allowlist validation and Deno sandboxing.

✅ Phase 2: Orchestration & Evaluation

  • Graph Topology: Natural language goal → DAG weaving via DSPy Module.
  • Wave Execution: Parallel async execution of independent node waves.
  • Triangulated Verification: Three-level evaluation (Rules → Semantic → HITL).

✅ Phase 3: Interactive Oversight (v0.2.5)

  • Goal Clarification: Intelligent pre-weave intake to resolve ambiguity.
  • Interactive Healing: Human-led failure diagnosis and repair guidance.
  • Final Approval Gates: Verification loops to ensure user satisfaction.
  • Documentation: Professional Diátaxis-structured engine docs.

🚧 Active Development (v0.3.0 - v0.5.0)

📍 Phase 4: Stability & Persistence [NEXT]

  • Wave-Level Checkpointing: Persistence of intermediate graph states to allow resume on crash.
  • Session Resume: Full CLI support for arachne resume <session-id>.
  • Semantic Topology Search: Replace SHA256 exact matching with vector-based fuzzy reuse of successful agent graphs.
  • Input Validation: Strict schema validation for node-to-node data passing.

📍 Phase 5: Observability & Streaming

  • Event Bus: SSE streaming of NodeStarted, TokenEmitted, and AutoHealTriggered events.
  • CLI Progress Streaming: Real-time visual feedback for long-running tasks.
  • Async Refactor: Eliminating nested asyncio.run calls for performance.

🔭 Long-Term Vision (v1.0.0+)

🛠️ Advanced Tooling

  • Playwright Stealth Agent: Autonomous browser interaction with anti-bot resilience.
  • Credential Vault: Encrypted JIT injection of secrets into agent modules.

🛡️ Security & Sandboxing

  • Secure Code Execution Environments: Implement fully isolated backend sandboxing (via Docker, E2B microVMs, or Deno) for safe evaluation of agent-generated Python and Javascript code.

🧠 Thick Intelligence Replacements

  • Framework-Level Automated Learning: Wire global memory tools directly into the core engine. Auto-healer writes resolved failure lessons to memory; Weaver pre-fetches memory to avoid repeating past graph mistakes.
  • Architecture Critique: Semantic review of the generated DAG before execution begins.
  • Just-in-Time Tool Broker: Dynamic discovery of tools based on runtime failures.

🏗️ Technical Debt (Future GitHub Issues)

These technical improvements are prioritized for backend maintenance:

  • Dependency Injection: Replace global Settings with constructor injection.
  • Service Abstraction: Create formal interfaces for MCPManager and SessionCoordinator.
  • Logging Standardization: Comprehensive migration to structlog across all modules.
  • Test Coverage: Reach >80% coverage for core execution modules.

🤝 Community & Support

  • Issues: Report bugs or request features via GitHub Issues.
  • Security: Report vulnerabilities to dan@strategicautomation.com.
  • License: MIT (See LICENSE for details).