Architect

Example worlds

GitS ships runnable worlds under architect/worlds/—each is a bundle of rules, archetypes, and (often) a matching YAML manifest in config/. Below: a system overview gallery from core dashboard flows, then two contrasting worlds with screenshots from runs.

System overview

Operator-facing surfaces: control plane, briefing, tasks, tools, knowledge, skill insights, events, LLM debug, per-agent token telemetry, and provider selection—plus integrations and secrets below.

Agent control plane view in the GitS dashboard — Agent control plane

Agent briefing panel with context and instructions — Agent briefing

Self task management UI — Task management

Tool usage detail view in the GitS dashboard — Tool usage

Knowledge entries visible to the agent — Knowledge view

Skill insights and skill-aware context in the dashboard — Skill insights

Per-agent token usage for a single agent — Agent token usage

LLM provider selection and model configuration — LLM providers

Integrations & configuration

First-class channel hooks, endpoint overrides, and secret handling from the same operator shell.

GitHub integration in the GitS dashboard — GitHub integration

Slack integration in the GitS dashboard — Slack integration

LLM endpoint override settings — LLM endpoint overrides

Secret management UI — Secret management

world0-Standalone-Complex

Standalone Complex

Single high-capability operator

A deliberately single-agent world: one “super-operator” (proto-9 / Proto) with broad tool and syscall access—channels, workspace, tasks, knowledge, and state—without a squad to delegate to. The name nods to Stand Alone Complex: emergent behavior without a central author; here that idea collapses into one agent that effectively is the network for that run.

Archetype: standalone-operative. Typical use: full-stack automation, recon, and long-running self-queued work when you want maximum leverage from a lone ghost.

Gallery

Code agent flow from a Standalone Complex run.

Standalone Complex code agent overview — Code agent

world6-Mafia-Town

Mafia Town

Multi-agent, conflict-forward demo — classic Mafia, moderated in-character

Demo: This is GitS as a multi-agent showcase with built-in conflict—hidden roles, opposing teams, and public pressure—so you can see channels, knowledge, and syscall.ask under tension, not only cooperative workflows.

A social game world: eight neutral player agents (town-1 … town-8) plus a dedicated moderator that runs phases, private role delivery, public narration, and endgame. Mafia vs Town roles are assigned at match start and live in moderated knowledge—not in agent ids—so the runtime stays symmetric until the story begins.

A gated Townfolks channel carries the public table; private flows use syscall.ask and knowledge entries for logs and team boards. Rules flow through prompts and moderator behavior rather than a separate Go rules engine—enough structure to play, flexible enough to improvise.

Archetypes: moderator and mafia-town-player.

This world’s Markdown, laws, and archetypes live in the open-source tree architect/worlds/world6-Mafia-Town.

Gallery

Multi-agent Mafia Town run at a glance.

Mafia Town multi-agent game overview — Multi-agent game

Source layout

Markdown and laws live in architect/worlds/<world-name>/ (e.g. world.md, laws.md, archetypes under archetype/). Boot-time roster and display names are wired through world config in the GitS repo on GitHub—see config/world0-standalone-complex.yaml and config/world6-mafia-config.yaml for these two.