How the request reaches the LLM — and who owns the loop

The defining decision each project makes: where does the model call live, and where does the agent loop live? Everything else — transport, persistence, auth — flows from that.

Pick a provider and step through the lifecycle. Each lit node is where bytes actually move; arrows that turn clay show the direction of the current step. The dashed box appears only when the architecture has a loop owned by that layer.

This is the most misunderstood part. "Keeping the loop going" means very different things in each system. The loop is the agent's autonomy — reasoning, calling tools, reasoning again until it decides it's done.

// opencode's explicit loop — the thing T3 deliberately does NOT have
const runLoop = Effect.fn(function*(sessionID) {
  let step = 0;
  while (true) {
    yield* status.set(sessionID, { type: "busy" });
    const msgs = yield* MessageV2.filterCompactedEffect(sessionID);
    const { lastAssistant, hasToolCalls } = latest(msgs);

    // EXIT: assistant finished AND not waiting on tool results
    if (lastAssistant?.finish && lastAssistant.finish !== "tool-calls" && !hasToolCalls)
      break;

    step++;
    const result = yield* handle.process({ messages, tools, model, ... });
    // ↑ streamText() → tool-call events → execute tools locally → write results
  }
});

Every cell below is grounded in source. "T3 (Claude)" and "T3 (Codex)" share the orchestration layer but differ at the adapter; they're shown as one column each because the contrast is instructive.

The headline finding from the source: T3 has no subscription system at all. So "how does the subscription create the T3 experience?" has a subtle answer — the subscription that matters is the one you hold with Anthropic or OpenAI, not with T3. Here are the three real models and what each unlocks.

What the experience layer actually delivers

Same prompt, same model — the "experience" is the orchestration around the call. Here's what each layer adds on top of raw model access.