mantis-colony
Queen and stigmergy. Both, not either.
Current multi-agent runtimes force a choice: hierarchical control (LangGraph, CrewAI) or stigmergic emergence (environment-signaled swarms). The first bottlenecks on the coordinator; the second degrades into noise at scale. Real colonies solve this by running both. A queen sets direction; workers broadcast TTL-decayed traces that other workers follow or ignore based on signal strength. mantis-colony ships this hybrid model as a single runtime, MCP-native at every layer, so any agent client can participate without protocol translation.
What mantis-colony can do
- Hybrid coordination: Queen daemon for high-level orchestration, stigmergic substrate for worker-to-worker signaling
- TTL-decayed memory traces with configurable half-life: strong signals propagate, stale signals decay to zero
- MCP-native daemon: every agent, signal, and coordination surface speaks Model Context Protocol
- Signal-strength routing: workers follow relevant traces, ignore weak or decayed ones, no central arbiter needed
- Multi-surface control: MCP daemon, CLI, web dashboard (desktop planned)
- Configurable decay constants: tune signal persistence per workload, from tight task graphs to long-lived research swarms
Queen daemon handles topological decisions (which agents exist, what work needs doing). Stigmergic layer handles operational coordination (which agent picks up which task, when to back off). Tuning the decay half-life between these two layers is the central design problem.
