reading · 7 min

Going Further

Objective: synthesise the curriculum and point to the real-integration next step and a capstone project.

Every module built one piece of the same machine: a loop that perceives and acts; a policy seam separating orchestration from the decision-maker; tools that let the agent reach the world; retrieval that grounds answers; memory that carries context across turns; planning for large goals; multi-agent coordination; evaluation; guardrails; a reliability layer of budgets, retries, and fallbacks; observability; and a stateless handler that composes it all for real traffic.

The through-line

The through-line across all twelve modules is the same idea from Foundations: the loop calls the policy, the policy returns a decision, the loop acts on it. Everything else is scaffolding that makes the core reliable, safe, and measurable. The loop does not change. What changes is how capable the policy becomes, how rich the context you give it, and how carefully you evaluate and guard its decisions.

The capstone: compose these for one real task

Pick one real task — document answering, data extraction, code review, or anything that matters to you — and compose the pieces you have built. Wire a real policy behind the seam. Add the tools the task needs. Plug in retrieval if it is knowledge-intensive. Apply budgets and guardrails appropriate to the risk. Instrument with traces. Run your evaluation harness against a golden set. Deploy behind a stateless handler.

You already know how each layer works. The capstone is integration: making them compose correctly for one specific problem.

Where real APIs and frameworks enter

Real model APIs and orchestration frameworks slot in at the policy seam (real integration: subsystem D). A framework is a pre-built implementation of the loop, the tool registry, and often the memory layer — it does not change what you need to evaluate, guard, or observe. SDK versions change, frameworks are deprecated; you can migrate without touching your evals or your guardrails because those live outside the implementation.

Best practice: frameworks change; the loop, the seam, evals, and guardrails do not — those are the job.

Durable principles outlast SDK churn

The principles in this curriculum are structural, not syntactic. Statelessness is a design decision. Evaluation is a discipline. Guardrails are a contract. Observability is an architecture — one event per step from day one, not a dashboard bolted on later. They outlast any particular SDK or provider.

You now have the full picture. The real work is building something with it.

Next: Production & Deployment Check