The quiet shift toward smaller, specialised agents

One assistant, one job

The shift is not driven by fashion. It is driven by the three things that keep enterprise platform leads awake at night — evaluation, cost, and the awkward question of who is accountable when the model gets it wrong. Narrow agents help with all three at once, and the compromises are easier to articulate to a board.

A team running eight focused agents can tell you, honestly, which two of them are saving money and which six are merely interesting. The same team running one general-purpose assistant usually cannot.

What the orchestrator actually looks like

We hand out a tiny scheduler, not a framework. Four callables, a typed registry, and a policy that can say no. Most of the interesting questions then happen at the edges, which is where you want them.

export async function route(
	req: AgentRequest,
	registry: AgentRegistry,
): Promise<AgentResult> {
	const plan = await policy.decide(req);
	if (plan.kind === "refuse") return { ok: false, reason: plan.reason };

	const agent = registry.resolve(plan.agentId);
	const result = await agent.run(req);
	await telemetry.record({ req, plan, result });
	return result;
}

What we are watching next

The interesting question for 2026 is not whether specialised agents work — they do ¹ — but what the connective tissue between them looks like when the list grows from eight to eighty ². That is the problem we are currently sitting with, and the answer has already stopped looking like a product and started looking like an operating model ³.