L1
Directive layer
Human-authored SOPs in Markdown. Versioned directives serve as the absolute source of truth for agent behavior.
Local-first autonomous runtime
AI-Tadpole-OS
A sovereign command layer for multi-agent work: Rust execution, React observability, private memory, and human-governed automation in one local runtime.
Runtime model
Built to keep strategy close to the machine.
AI-Tadpole-OS treats a clone as the unit of sovereignty. Directives, orchestration, and execution remain inspectable so teams can run missions without surrendering operational context to a hosted black box.
L2
Orchestration layer
Agent 99 coordinates swarm hierarchy, decomposes objectives, and performs Self-Annealing to refine long-term memory.
L3
Execution layer
Rust-native engine and deterministic scripts execute tools in parallel using FuturesUnordered for absolute speed and security.
Operator interface
Dense control surfaces for live swarm work.
The product codebase exposes a high-density dashboard system rather than a chat-only wrapper. Operators can inspect agents, model slots, quotas, telemetry, mission state, and detached tactical views from the same shell.
- Operations dashboard with 10Hz Swarm Pulse (MessagePack telemetry) and detachable portals
- Model manager with triple-slot routing (Primary, Secondary, Tertiary fallback slots)
- Codebase Knowledge Graph HUD with dynamic traversal history and BFS Dependency Pathfinder
- Security dashboard featuring the Sapphire Shield and OBLITERATUS Merkle audit trails
Model intelligence
Provider choice without operational drift.
The model manager is designed around capability-aware routing: local/self-hosted models, hosted providers, secure vault sync, modality filters, and triple-slot fallback per agent.
Providers
Ollama, OpenAI, Groq, Google, and custom nodes
Routing
Primary, secondary, and tertiary model slots per agent
Industry templates and MCP skills are first-class operational material, not marketing add-ons. They feed directly into the agent, workflow, and capability model.
Governance
Autonomy with explicit control points.
AI-Tadpole-OS is built around bounded execution. Sensitive actions can require human approval, mission spending is visible, and local-first operation can be enforced when privacy is the priority.
Hard Privacy gate
Block external cloud traffic entirely when a mission requires 100% air-gapped isolation.
Sapphire Shield
Mandatory human gating for high-risk tool calls like budget:spend and shell:execute.
OBLITERATUS Hardening
Audit-verified code paths, zero context decay, and Merkle-proof decision history.
Secret-aware logs
Real-time token and credential scrubbing before telemetry reaches LanceDB storage.
Roadmap
From local engine to distributed swarm mesh.
The May 1, 2026 product roadmap shows the platform moving through governance, memory, distributed runtime, and hardened infrastructure work.
Current
Rust backend, reactive dashboard, modular skills, model registry, local-first memory
Next
Continuity scheduler, unified oversight gate, GraphRAG, communication bridges
Distributed
Bunker mesh, portable context cores, graceful provider degradation, self-annealing recovery
Clone the runtime
Own the directives, memory, and execution path.
Start from the source repository, review the architecture, then run the Rust engine and React dashboard locally.