modiqo/waggle

What it solves

Waggle addresses the inefficiency and lack of accountability in multi-agent handoffs. Currently, orchestrators typically forward entire context blocks to subagents, which leads to high token costs, context window saturation, and a lack of telemetry regarding whether a subagent actually read the required information. It replaces the practice of passing raw file paths (which lack versioning and attribution) or full text (which is expensive) with a lightweight, attributed reference system.

How it works

Waggle uses a ~30-byte token as a reference to an artifact. Instead of sending the artifact itself, the orchestrator sends the token. When a subagent resolves the token, it receives a "projection"—a version of the data tailored to its specific model or needs (e.g., a summary or a specific slice) under a byte budget.

Key technical components include:

• MCP Server: It integrates as a Model Context Protocol (MCP) server, allowing it to be added to tools like Claude Code, Cursor, and Codex without custom SDKs.
• Attribution Manifest: Tracks who minted the token, the lineage of delegation, and variants for different consumers.
• Event Log: An append-only log records every read and resolution, providing a verifiable audit trail (receipts) of what was accessed.
• Lenses: Specialized engines for different file types; for example, source code is processed with tree-sitter at mint-time to allow agents to read specific symbols without parsing the whole file during retrieval.
• Edge Deployment: Tokens can be federated across machines or pushed to Cloudflare's edge, allowing remote agents to search or resolve content without the source files ever leaving the original host.

Who it’s for

It is designed for developers building multi-agent orchestrators and harnesses (such as those using Claude Code or Codex) who need to reduce token overhead, ensure subagents are actually consuming the correct data, and maintain a consistent reference layer across different AI vendors and environments.

Highlights

• Token-based Handoffs: Replaces massive context dumps with 30-byte references.
• Verifiable Consumption: Provides "receipts" and coverage reports to prove which parts of an artifact were read by an agent.
• Context-Aware Projections: Delivers only the necessary slices of data to the consumer based on their context and budget.
• MCP Integration: Works out-of-the-box with any MCP-compatible AI harness.
• Symbol-Level Access: Enables agents to target specific code symbols directly via tree-sitter outlines.
• Distributed Reach: Supports local, federated, and edge-based resolution of tokens.