The loop is the architecture's heart

Every agentic system, whatever framework built it, runs the same cycle: given a goal and the story so far, the model proposes a step; a tool executes it; the result comes back as observation; the model decides again. Single-pass generative systems answer and stop — the loop is what lets behaviour accumulate across steps, and it is also what lets errors accumulate, which is why the loop's exit conditions (step budgets, cost ceilings, terminal states) are first-class architecture rather than afterthoughts.

Four parts, whatever the framework

Strip the implementation and four components remain. The model proposes next steps. Tools — increasingly exposed through the Model Context Protocol — are the named, scoped capabilities it may invoke, and their list is the system's true permission boundary. State carries what has been tried and learned across steps, from a scratchpad in context to durable stores for long-running work. Orchestration runs the loop: sequencing, retries, budgets, hand-offs between multiple agents, and the human-approval pauses on consequential actions.

The patterns you will meet

Recurring design patterns compose the same parts differently. Reflection adds a verify step where the model critiques its own output before acting on it. Planner–executor splits the roles: one pass drafts the plan, another carries it out. Multi-agent designs give specialised agents their own loops and route work between them — which multiplies capability and also multiplies the surfaces to govern. Human-in-the-loop is a pattern too, not a bolt-on: the approval gate is part of the architecture, placed where actions become hard to reverse.

Architecture is where readiness is decided

Most operational properties are set by these choices, not by the model. Whether an incident can be reconstructed depends on whether every loop step was traced. Whether an agent can be contained depends on whether tools carry their own scoped credentials. Whether autonomy can widen safely depends on whether the orchestrator enforces evaluation-gated levels. The build guide walks these decisions in order — and the pillar page puts the architecture in its wider context.