agentglue 0.4.0

Thin runtime layer for multi-agent systems — cross-process dedup, TTL cache, single-flight, SQLite backend, and observability

AgentGlue

Middleware coordination for multi-agent systems.

Hilights • Key features • How it works • Install • Quick example

AgentGlue is a middleware layer that sits between your orchestrator and your tools, helping multi-agent systems coordinate overlapping work instead of repeating the same reads, searches, API calls, and state lookups over and over again.

📈 Hilights

Measured on the current 100-test benchmark suite and overlap-heavy multi-agent workloads:

• 3.7× overall speedup across benchmarked scenarios
• 73% total time saved (866.4ms → 234.9ms)
• 76% cache hit rate across repeated shared work
• 6.8× speedup on overlapping search-heavy scenarios
• 5.0× speedup in the 10-agent heavy-overlap case
• ~0.6ms median cache-check latency

More agents + more overlap = bigger wins.

✨ Key features

• Exact-match dedup on tool name + args/kwargs hash
• Single-flight / in-flight coalescing for concurrent identical calls
• TTL result cache for repeated sequential calls
• SQLite-backed cross-process cache for shared coordination across processes
• Metrics and event recording for runtime inspection
• Keep shared work coordinated across agents, tools, and processes
• Stay framework-agnostic across OpenClaw, AutoGen, CrewAI, LangGraph, and custom systems

🧭 How it works

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃  🤖 Agent A    🤖 Agent B    🎯 Orchestrator       ┃  ← Intelligence Layer
┃                                                   ┃    Autonomous reasoning
┃  "Decide what to do"                              ┃    & task decomposition
┣━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┫
┃                 ⚙️  AgentGlue                     ┃  ← Coordination Layer
┃         (Coordination Middleware)                 ┃    The "glue" that binds
┃                                                   ┃    agents to resources
┃  ┌──────────┬──────────┬──────────┬──────────┐    ┃
┃  │ Dedup &  │  Rate    │ Task     │   Shared │    ┃    · Dedup & coalescing
┃  │coalescing│  Mgr     │ Locking  │   Cache  │    ┃    · Rate coordination
┃  └──────────┴──────────┴──────────┴──────────┘    ┃    · Task Locking
┃                                                   ┃    · Shared Cache
┃  "Know how to coordinate"                         ┃
┣━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┫
┃  🔧 Tools    📡 APIs    📁 Files    💾 Memory      ┃  ← Resource Layer
┃                                                   ┃    External capabilities
┃  "Do the actual work"                             ┃    & persistent storage
┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛

AgentGlue sits in the middle: agents keep their existing orchestration logic, tools keep their existing interfaces, and AgentGlue coordinates repeated work across both.

🛠 Install

1) Install the OpenClaw plugin

openclaw plugins install openclaw-agentglue

After install, restart the gateway:

systemctl --user restart openclaw-gateway

If you want to remove the plugins.allow security warning, add this to ~/.openclaw/openclaw.json:

{
  "plugins": {
    "allow": ["openclaw-agentglue"]
  }
}

Plugin docs: openclaw-agentglue/README.md

2) Install the Python package

From PyPI:

pip install agentglue

Or from source:

python3 -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -e .

⚡ Quick example

from agentglue import AgentGlue

glue = AgentGlue(shared_memory=False, rate_limiter=False, task_lock=False)

@glue.tool(ttl=60)
def search_code(query: str) -> str:
    return f"results for {query}"

print(search_code("sqlite sidecar", agent_id="agent-a"))
print(search_code("sqlite sidecar", agent_id="agent-b"))
print(glue.report())

🔌 Links

• GitHub releases: https://github.com/brickee/AgentGlue/releases
• OpenClaw plugin: https://www.npmjs.com/package/openclaw-agentglue
• Plugin docs: openclaw-agentglue/README.md

⚖️ Disclaimer

This project is a personal open-source project developed in authors' personal capacity. It is not affiliated with, endorsed by, or representing any employer or organization the authors are associated with. All work on this project is performed on personal time and with personal resources.

📄 License

MIT