Live technical snapshot.

This is the layer that decides which model gets which job. It should route by task type, risk, speed, cost, and fallback rules instead of hard-coding one favorite model everywhere.

• Route by task, not by brand.
• Keep a fallback path when a provider is down or no longer worth the cost.
• Keep routing policy separate from prompt text so it can change cleanly.

This is how work moves through the system: plan, execute, review, approve, publish. It includes agent roles, task handoffs, retries, and the small reusable behaviors around them.

• Separate planning, execution, and review.
• Use clear task contracts, not vague prompts passed from step to step.
• Keep skills, hooks, and automation around repeatable points in the workflow.

Models are only as good as the context they get. This layer keeps prompts, docs, retrieval, decisions, and working state useful without stuffing everything into one giant prompt.

• Keep context scoped to the job.
• Store durable knowledge in docs, indexes, and retrieval systems.
• Reset stale state instead of letting it pile up.

This is where the system touches the outside world: files, APIs, databases, search, GitHub, terminals, and optional protocol adapters.

• Prefer simple local interfaces first.
• Use MCP or similar adapters when they reduce integration work, not as the foundation of the stack.
• Keep tool access explicit so it is easy to audit and replace.

You need a way to know whether the system is improving or drifting. This layer covers evals, observability, rollout rules, and team adoption.

• Use eval sets and approval gates before high-impact changes.
• Track cost, quality, and failures in the open.
• Make rollout, rollback, and team training part of the stack, not an afterthought.