Fourteen capabilities. One runtime.
Each one has its own deep dive, sourced from the live codebase docs, with the real admin screenshot where one exists. Click in.
Memory
Four memory layers, retrieved selectively, consolidated nightly.
Execution layerTools
35+ built-in tools. Unlimited MCP servers. A nine-point safety gate every call.
Knowledge layerContext
Five-layer context budget. Four-stage compaction. The model never runs out of room.
Control layerWorkflows
Markdown SOPs that cascade through products, goals, projects, and tasks.
Control layerTask management
Tasks, projects, goals. Prioritized, assigned, picked up by the right agent.
Control layerHeartbeats
Assess. Then execute. Recoverable, inspectable, on schedule.
Execution layerLocal models
Your GPU stays sane. User chat jumps the queue.
Observability layerObservability
A flame graph for agent execution. Failure taxonomy. Replay.
Execution layerMulti-provider
Native SDK per provider. Switch per agent, per task, or globally.
Knowledge layerSkills
Reusable methodology bundles. Discoverable, activatable, security-scanned.
Observability layerEvals
A/B comparison for agents. Across skills, across models.
Control layerApprovals
Three-tier sign-off. Versioned. Reversible. Audit-logged.
Observability layerCosts & budgets
Per-agent and per-tool spend. Threshold alerts. No surprises.
Control layerProduct OS
Twelve versioned plan documents per product. Agents author, you approve.