Defines the agent's identity and fundamental runtime behaviour. interaction_mode is the most critical field — it determines whether the agent converses or treats every input as a task.

Two-stage classifier that runs before decomposition. Stage 1 is a fast heuristic; Stage 2 calls an LLM only when Stage 1 is unsure. The Capability Scout pre-probes MCP tool servers to map capabilities that steer decomposition.

Defines the agent's declared capabilities — what kinds of work it knows how to do. The decomposer matches user requests to task types and dispatches sub-agents. Each entry has its own tool routing, timeout, retry policy, and optional blueprint guidance.

MCP servers that sub-agents can call during task execution. Each entry is a named connection — local or remote — with its own auth, TLS, pooling, and health check settings. These are the MCPs you explicitly configure and trust.

Defines the LLM backends available to the framework. The default provider is used by the primary agent and any tasks that don't specify their own. Named providers allow different parts of the framework to use different models.

Per-task guidance documents accumulated from real execution runs. Each blueprint captures topology hints, dos/don'ts, known failure modes, and lessons learned. When present for a task, it is injected into the primary agent's system prompt before decomposition.

When a task has no matching tool server, the agent generates Python scripts on the fly and runs them in an isolated sandbox. This registry is the index of all scripts the agent has created — their metadata, use counts, and whether they've been promoted to permanent handlers.

MCP servers discovered automatically by the Capability Scout from public registries. Unlike Tool Servers you configure manually, these are discovered at runtime. Each carries a trust tier, auth requirement flag, and verification timestamp.

Self-spawning specialist agent mesh. When the primary agent encounters a capability gap, it can hatch a new "ant" — a lightweight Cortex instance pre-configured for a specific domain. Each ant runs as a separate process on its own port and registers itself as an MCP server.

Covers the autonomic learning engine, session validation, history retention, storage backend, security guardrails, startup behaviour, and the Cortex UI. Also shows the live learning delta queue with proposals waiting for confidence-based promotion.