What the engine does for you
Every agent you build runs on the Nexus engine, which enforces a set of behaviours on top of whatever you configure. You don't switch any of these on - they apply to every agent, single and multi alike.
This page is worth reading once before you write instructions, because the most common cause of a bloated prompt is defending against something the platform already handles. You don't need a rule telling the agent to stay on topic mid-flow, and you don't need to explain how to recover from a failed tool call.
| Area | What the engine does | Why it matters |
|---|---|---|
| Routing | Two stages - a small Context Expert picks the agent from each agent's Trigger alone, then the chosen agent runs with its own instructions and tools. | Faster routing, a narrower second prompt, fewer wrong-tool calls. |
| Continuity | A fixed rule every turn: stay on the same agent unless the user has clearly changed topic. | Fewer mid-flow drop-offs, without writing defensive "stay here" instructions. |
| Numbered steps | Instructions that list Step 1: / Step 2: … get sequential-execution rules injected automatically - one step per turn, with progress tracked. | Multi-step procedures stay on track without boilerplate. |
| Conversation history | Past roughly 15 messages or 10K characters, older turns are summarised automatically; the last two user turns stay verbatim. | Long conversations don't exhaust context, and recent detail survives. |
| Tool failure | The same tool is never silently retried - the error comes back as the tool result so the agent can phrase a recovery. | Predictable failure behaviour you can write instructions against. |
| Execution trace | Every turn stores a structured trace: tool calls and results, memory writes, and the routing decision. | The Testing lab and debugging read the same record, so a bad production conversation becomes a regression test in one click. |
| Versioned configuration | Every Configuration change is versioned, with history. | Roll back, and review how behaviour shifted over time. |
@-mentions in prompts | Type @ in most prompt fields to insert a live reference to an agent, workflow, tool or rich-media item. Chips survive renames. | Prompts don't go stale when you rename their target. |
Where each one is covered in depth
| Behaviour | Full detail |
|---|---|
| Routing and the Context Expert | Routing logic |
| History compaction and what agents share | Memory |
| Traces and turn-by-turn debugging | Traces |
Next: Configuration - the settings that shape every agent.