Orchestration
Multi-agent collaboration
Isolated workbenches with built-in peer review. Parallel execution across tasks. Pluggable agent seats assigned by capability.
Overview
Core ideas
State
Stages & runtime truth
Every task moves through defined stages with explicit transitions. One canonical state file per task tracks stage, round, active agent, and pending approvals.
Intelligence
Context engineering
Prompts assembled from live runtime state and policy constraints rather than static templates. Optional Cortex retrieval augments intake when configured.
Fitness
Qualification registry
Measured fitness per agent per task type, recorded in an evidence-backed registry with explain surfaces. The registry emits advisory warnings today; automatic assignment changes are planned, not shipped.
Integration
Dev infrastructure
Git workflows, PR management, workspace orchestration, and bench coordination. The runtime connects to the tools and processes that real multi-agent development depends on.
Adaptivity
Adaptivity
Pluggable adapters for different AI engines. Configurable seats, policies, and automation levels. The runtime adjusts to what's available and what works.
Design
How work gets done
- Design before code — write down what should be true and what would prove it wrong before implementation starts
- Staged work cycles — intake, planning, implementation, verification, adversarial review, closeout
- Multi-perspective analysis — parallel agents explore from different angles, cross-check, and converge on a result
- Built-in peer review — one agent works, another reviews with reject/fix cycles
- Automated consistency checks — cross-references, timestamps, and artifacts validated at each stage boundary
- Structured prompts — every instruction built from current state, policy constraints, and the specific task at hand
- Fitness measurement — track which agents do well at which tasks and use that evidence for future assignment
Development
What we're building
- Feature data pending.
Explore
Links
Requirements
Dependencies
- bash 5.0+ — runtime shell
- git 2.30+ — worktree isolation and branch management
- tmux 3.2+ — agent session hosting and seat panes
- Node.js — launcher and operator view surfaces
- Agent CLI — Claude Code, Codex, or any engine with a CLI prompt interface
- Cortex (optional) — prompt intelligence service for retrieval-augmented guidance
Architecture
Local-first, hostable, measurable
Runs on your machine. Manages agent sessions through tmux. State in plain files. Prompts assembled from retrieval and live context. Fitness measured and fed back. Projects to hosted views when ready.