Harness - build AI agents
Harness is where you build AI agents on Nexus. You design one or more specialised agents, shape how all of them behave with a shared Configuration, and give them tools to do real work.
This page is the map - what the pieces are, and where each one lives.
The three pieces
Agents
Agents are the specialists, each designed for one job - an order-status agent that answers "Where's my order?", a demo-booking agent that helps someone pick a slot, a billing agent that handles refunds and invoices. They come in two types:
- Single agents - one agent, one prompt. It reads the conversation and decides each turn what to ask, which tool to call, and when the job is done. You describe the outcome; the model finds the path. Best when the path genuinely varies: support, troubleshooting, open-ended Q&A.
- Multi agents - the same reasoning, split across steps you lay out on a canvas. Each step is small and focused, and you draw the routes between them. Best when the procedure matters as much as the answer: order returns, KYC checks, loan applications.
Both use an LLM. The difference is who owns the path - a single agent works it out, a multi agent moves along routes you defined. Inside a multi agent, the model still runs the conversation at each step and picks which exit it takes; what it can't do is invent a step you never defined, or skip one you did.
That's also why the more complex work tends to land on multi agents. A single prompt only carries so much before instructions start getting dropped - a twenty-step procedure with validation, branching and rollback doesn't fit in one reliably. Splitting it into nodes keeps each prompt small, and makes a run inspectable step by step instead of one opaque decision.
You create both from the same Create picker - start at Create an agent, then continue to single agents or multi agents.
Routing decides which agent takes a given message. The primary signal is each agent's Trigger - the natural-language description of when it should take over - and you can add Routing logic rules on top when Triggers alone aren't precise enough. Once an agent takes over, it owns the conversation until control returns.
Configuration
Configuration is project-level: what you set here applies to every agent, single and multi alike. Identity and tone, conversation rules, routing logic, global tools, memory, fallback, lifecycle hooks, and AI safety all live here.
That makes it the highest-leverage surface in Harness - one change shapes every conversation. See Configuration for the full tour.
Voice settings is also project-level, but lives on its own surface under AI Agents → Voice rather than inside Configuration.
Tools
Tools are the actions an agent can take in the real world: call a workflow, search a knowledge base, escalate to a human, transfer a call. You define a tool once and any agent can use it - and it can be called at the Configuration level too.
Attach a tool to a specific agent when it belongs to that agent's job. That also helps route the topic to the right place. See Tools.
What the platform does for you
The Nexus engine enforces a set of behaviours on top of whatever you build - two-stage routing, automatic history compaction, parent-child agent calls, structured traces - so you don't build them yourself.
You don't configure any of it. See What the engine does for you for the full list, and read it before you write defensive instructions for something the platform already handles.
Where each surface lives
The product calls this section Harness. It has three groups in the left nav:
| Group | Items |
|---|---|
| AI Agents | Agents (single and multi) · Configuration · Widgets · Tools · Voice |
| Automation | Flows · Background workflows · Functions · API · Scheduler · Localization |
| AI Trust Center | Overview · Testing lab · Evaluators & rules · User profiles · Action center |
Note: what you see depends on your role and your setup - Background workflows and API in particular. If an item in this table isn't in your nav, that's why.
Where to go next
By what you're trying to do, rather than in reading order:
| Build something now | Quickstart - one agent, one tool, one test, in about 15 minutes |
| Set behaviour for every agent | Configuration - identity, rules, routing, memory, fallback, safety |
| Design an agent | Create an agent → single agents or multi agents |
| Give agents something to do | Tools - workflows, knowledge base, escalation |
| Lay a procedure out step by step yourself | Multi agents |
| Add custom UI to a conversation | Widgets |
| Take it to phone calls | Voice |
| Know it still works tomorrow | Testing and the AI Trust Centre |
| Follow a whole build end to end | Guides - real outcomes, crossing whatever surfaces they need |
Not sure whether you need one agent or several? Start with one. A single agent with clear instructions handles more than people expect, and routing only earns its keep once you have genuinely separable tasks. Create an agent routes the common jobs to a type.