Collapse Model Hops — Don't Chain Specialists When One Model Can Do Both
Every model hop in a multi-model pipeline is a place where context gets translated, latency stacks, and errors propagate. Collapse perception and action into a single call when the model supports it.
A common agent architecture chains specialist models: a vision model reads a screen, passes a text description to a reasoning model, which passes a plan to an action model. Each hop is a place where context gets translated (losing fidelity), latency stacks (each call adds round-trip time), and errors propagate (a misread in hop 1 becomes a wrong plan in hop 2).
When a single model can handle multiple modalities and capabilities natively, collapse the hops. A model with native vision, search grounding, and tool use in one call can look at a screen, ground a fact against live search, and pull a location — without handing context between three specialized models.
- Computer-use agents: A model with native screen understanding + action generation (e.g., Gemini 3.5 Flash with computer use) replaces a vision-OCR → text-reasoning → action-execution chain
- RAG + reasoning: A model with native search grounding + reasoning (e.g., Gemini with Google Search grounding) replaces a retrieve → summarize → reason chain
- Multimodal analysis: A model with native image + text input replaces an image-caption → text-analysis pipeline
- Tool use + planning: A model with native function calling + reasoning replaces a planner → executor split when the task doesn't require them to be separate
- Cost separation: If the specialist model is dramatically cheaper for its sub-task, chaining may be more cost-effective than calling a larger multimodal model for everything
- Quality separation: If a specialist model is meaningfully better at its sub-task than the generalist, keep the hop — the fidelity gain is worth the latency
- Independent scaling: If different hops have different load patterns, separate models let you scale them independently
- Vendor lock-in risk: Collapsing into one model's native capabilities ties you to that model's feature set
Collapsing hops trades flexibility for reliability. Fewer moving parts means fewer failure points, lower latency, and less context loss — but it ties your agent's capability surface to what one model supports natively. The right answer is usually: collapse where a single model is good enough at both sub-tasks, keep the hop where a specialist clearly outperforms.
Models shipping in September 2026 are converging on this. Gemini 3.5 Flash ships computer use, search grounding, and tool use in a single call. Claude Opus 5.5 targets long-running agentic coding where the same context persists across many turns. The trend is toward fewer, more capable calls rather than more, narrower ones.