Building a Mixed-Adapter Team

Building a Mixed-Adapter Team

Different agents in your company should not all run on the same adapter by default. Adapter choice should follow the role's demands — and different roles make very different demands. Getting this right from the start saves money, produces better output, and avoids adapter mismatch bugs that are frustrating to diagnose.

---

Why adapter choice matters per-role

Every adapter has a different cost profile, capability set, and operating model. Running every agent on your personal favourite adapter is a reasonable shortcut when you have one or two agents, but it becomes expensive and limiting as the company grows.

| Adapter | Strength | Cost profile | |---|---|---| | claudelocal (Opus) | Strategic reasoning, long context, complex writing | Higher — best reserved for roles that need it | | claudelocal (Sonnet) | Balanced reasoning, coding, writing | Medium — good general-purpose choice | | codexlocal | Code generation, editing, test writing | Comparable to Sonnet; optimised for code tasks | | geminilocal | Google ecosystem integration, multimodal inputs | Varies by model tier | | opencodelocal | Multi-provider flexibility | Depends on model selected | | hermeslocal | Persistent memory, research, knowledge accumulation | Medium — memory ops add cost; stateless tasks cost more than necessary |

The principle is simple: assign the adapter that fits the work, not the adapter you are most comfortable with. A Codex-based engineer costs less and performs better on code tasks than an Opus-based engineer would. An Opus-based CEO costs more but produces better strategy than a Sonnet-based one would. The match matters.

---

The principle: match adapter to role demands

Before hiring anyone, think about each role in terms of three questions: