With ATT (AI-Team-Team), AI can freely form teams, define how they interact, how teams collaborate, and create all sorts of incredibly complex hierarchical (or dynamic) relationships.

If you find this project interesting, please support it with a ⭐️ Star! Thank you so very much! 🙏

👉 Project Repository：https://github.com/AI-Team-Team/AI-Team-Team

🧬 Topology & Lineage Control

• Tree-like Lineage Spawning: Spawns recursive child agent teams (AgentTeam) at runtime to arbitrary depths, strictly bounded by depth limits to prevent stack overflow.
• Autonomous Member Configs: Defines dynamic child memberships mapping role presets, custom system instructions, and LLM aliases to shape custom agent personalities.
• Dynamic Lineage Migration: Permits active teams to request parent-hierarchy migrations, arbitrated by modular strategies with loop/cycle detection and parent notification logs.
• Hierarchical Topology Map: Injects an ASCII-drawn indented tree map of active teams (displaying purposes, status, and progress metrics in real-time) directly into the agent prompt context.
• Global Expert Discovery: Automatically appends a directory of all active system experts (names, roles, and profiles) into the agent's identity context to facilitate peer discovery.

🧠 ReAct Loops & Execution Engine

• Bounded ReAct Loops: Executes standard Thought/Action/Observation reasoning cycles, capped by max steps to prevent runaway API tokens.
• Robust Argument Parser: A safe literal lexical parser (ast.literal_eval) with multiline XML support, code block stripping, and a comma-merging heuristic to handle unquoted complex strings (like SQL queries).
• Dialogue Memory Compression: Automates memory pruning by summarizing early conversation turns using the agent's LLM while preserving the latest high-fidelity messages.
• Model Registry & Callbacks: Delegates all LLM generation logic to a single global callback handler (generator_handler), keeping the framework lightweight and vendor-independent.

🗳️ Governance & Inter-Team Communication

• Democratic Voting System: Features an asynchronous voting pipeline to add or remove members, requiring unanimous participation and a $\ge 2/3$ agreement majority.
• Anonymous Voting: Enforces voter anonymity via cast_vote(..., public=False) which masks voter names as "Anonymous Voter" in the team prompt context.
• Negotiation Broker: Regulates P2P messaging and agreement tunnels through dynamic parent rules (allowing regex purpose/lineage checks) or interactive LLM leader proxies.
• Collaborative DocLib Storage: Equips teams with built-in document libraries. Access is governed by prefix path ACL permissions (READ/WRITE) that inherit recursively downward to subdirectories.

🔒 Context Protection & Safety Gates

• Gated Context Protection: Restricts direct large file reads; falls back to Outline Warnings with a 5-line sample of files exceeding 50 KB, prompting agents to make paginated, sliced chunk requests.
• Tool Auditor Interception: Registers pre-execution interception hooks to audit, vet, approve, or reject specific tool calls (e.g. database safety query check).

💾 Persistence & Diagnostics

• SQLite State Snapshots: Automatically serializes topologies, lineages, memory logs, DocLib files, and active voting proposals to SQLite on state changes, enabling crash-recovery.
• Supervisory Dialogue Audits: A non-participating 3-AI Supervisory Team (Integrity, Continuity, Deadlock) reviews round transcripts, recursively escalating anomalies up the tree lineage.
• Decoupled Dashboards: Exposes clear runtime callback event hooks (on_status_change, on_activity_added, on_log_append) to update console UIs without codebase pollution.

(All the features mentioned above have been substantially implemented.)

🛠️ Current Status & Future Plans:

The project is currently a fundamentally complete Python package, through which you can observe some fascinating AI sociological behaviors (it has already helped me brainstorm quite a few great ideas). In the near future, I might launch a simple application for end-users based on ATT.

Many details are still rough around the edges, and some ideas are still being rolled out step by step. However, I personally believe the core concept is highly intriguing, and I warmly welcome your feedback, suggestions, and contributions 👍

Thanks for reading, and have a great one! Bye!