Dev prototypes 10 ways to navigate a 10,000-agent swarm with Mantiq observability tools

strickvl · x · 2026-10-31

Developer strickvl shared Mantiq, a toolset for AI engineers to work with agent swarm traces, plus ten prototype views for navigating a 10,000-agent swarm.

The problem: agents' own accounts of what they did may be wrong, and swarm failure modes are scattered, so multiple signals must be combined. Mantiq records five things: model and tool calls, every kernel process, every file read/write, inter-agent messages, and container output.

Notable views:

Constellation view: messages drawn as lines between agents/teams
Problem funnel (Sankey): what went wrong → the pattern behind it (e.g. parser.py never landed after 8 pushes) → judge scores; analysis auto-generated by the 'hodhod' agent
Stack view: 3D stack with teams on the floor and time going up; failures shared across teams appear as a layer, click any slice to inspect an agent's state
Treemap: size = work, color = failure mode
Drill-down from swarm → team → agent → individual message or file edit

He also designed a compact visual symbol language (took a task / never released / push refused / never collected, etc.) to avoid text-heavy UIs. Demo run: 1 planner + 16 workers + 1 judge on gpt-6-luna wrote a Python Lisp interpreter in 30 min for $0.18, passing 10/180.

Original post →

Dev prototypes 10 ways to navigate a 10,000-agent swarm with Mantiq observability tools

More from coding & agent