Agent Costs Are Rising — Composition Is the Answer

The default pattern for integrating enterprise data into AI is bolting tools onto general-purpose agents via MCP and skills. This inheritance model works for a handful of integrations but degrades as context bloat increases, creating fragile and expensive agents. Most teams haven't hit the ceiling yet, but token cost trajectories suggest they will.

Domain-specific agents propose composition: many small, highly focused agents with minimal context, each owning a narrow capability, coordinated by a lightweight orchestrator. The approach mirrors how human teams organize complex work. Early results from the speaker's stealth company claim over 80% token efficiency gains and the ability to use far cheaper models for sub-tasks.

The operational tradeoff is complexity. Managing many agents introduces coordination overhead, new failure modes, and requires strong telemetry and agent portability. The ecosystem currently lacks standards, making adoption risky. However, the reversal of the “cheaper intelligence” trend — token costs up 76% in 2026 — adds urgency for large-scale deployments.

Engineering leaders should treat this as an architectural bet, not a near-term playbook. Monitor emerging frameworks, enforce strict capability limits on any customer-facing agents, and start identifying narrow domains where a small specialist agent could replace expensive general-purpose inference. The prediction of late-2026 traction is plausible but unproven.