Experimental system
One workspace. Many agents.
Specialist agents, a shared model gateway, and inference distributed across local hardware.
Understand which agents run and where delays occur.
Specialist agents → shared gateway → local hardware.
Visibility into routing, missing traces and model loading.
Experimental system · No controlled task-quality benchmark published.
The problem
An agent team needs more than assigned roles. It needs model routing and visibility into which agents actually run.
My contribution
I assembled a local team of named agents and routed their model requests through a single gateway. I added infrastructure monitoring and execution traces to understand the system’s behaviour.
Architecture choices
Specialist roles
Zeus plans; Iris researches; Hermes supports sales; Echo handles marketing. Hestia, Atlas, Argus and Plutus cover support, coding, testing and finance.
Shared gateway
Model requests route to a DGX Spark, AMD Strix Halo box, Mac Mini or A4000 workstation.
Two views of execution
Grafana tracks hardware and model residency. Langfuse tracks agent calls, tokens and responses.
What it showed
Routing visibility
Tracing revealed which agents were being called and where token or timing data was missing.
Operational insight
Monitoring connected repeated model loading with slow task execution.
Status & next evaluation
Experimental system as described on 16 August 2026. Tool calling, mobile approvals and automatic output grading were listed as next steps in that article.
The next evaluation should test tool use, approval handling and output quality against repeatable tasks.
Project source
The project details and reported results come from my linked LinkedIn publication.
Read the architecture and monitoring article