No orchestrator, no MCP: 4 peer agents on a gossip mesh research the web and deliver to Slack

Trainer_Intelligent · reddit · 2026-09-04

A developer open-sourced jarviscore, a harness for running local AI agents as a fleet of equal peers — no orchestrator process, no MCP server. Four Python processes discover each other via a SWIM gossip mesh (the protocol HashiCorp uses for cluster membership) and claim tasks atomically from a shared Redis ledger; the demo shows four agents doing real web research through a local SearXNG and delivering results to Slack.

Key design points:

No single point of failure: kill any node and another claims its work
Claimed, not assigned: steps live in a Redis ledger with dependency gating on ledger state
No MCP server needed: 1,224 typed atoms across 150 services ship in the box; agents can also write sandboxed code, repair it, and graduate it into a verified registry
Zero-trust credentials: agents never hold keys; tokens are encrypted at rest and resolved at the call boundary by a broker, so a leaked trace leaks nothing
Fully auditable: each agent writes a flight recorder; one command replays the run turn by turn (176k tokens in the demo, all auditable)
Crash-safe: steps are checkpointed; rerunning the same workflow id restores completed work instead of re-running and re-billing

Open source via pip install jarviscore-framework, with runnable demos.

Original post →

No orchestrator, no MCP: 4 peer agents on a gossip mesh research the web and deliver to Slack

More from coding & agent