Agent Harness

What your agent does at runtime — how a turn behaves, the limits you can set, and the reliability, memory, orchestration, observability, safety, and evaluation features you get on every channel.

Everything you build in the dashboard runs the same way on every channel — web, widget, instant messaging, email, A2A, the OpenAI-compatible API, schedules, and webhooks. This group documents the runtime behavior you can rely on and the controls you can set, so you know what to expect when you put real traffic on your agent.

The builder pages tell you what to configure. This group describes how your agent behaves once it’s running.

What you get

Concern	What it means for you	Details
Execution	Every reply is a bounded turn with limits you can set on time, tool use, and spend	How a turn executes
Durability	Conversations, schedules, and queued work survive restarts and updates	Durability and recovery
State	Each channel keeps its own conversation; workflow progress and configuration history are tracked	State and sessions
Tool reliability	Approved actions aren’t repeated on a retry, calls are time-bounded, and destructive actions need consent	Tool execution and reliability
Memory & context	Long-term memory, knowledge retrieval, and automatic handling of long conversations	Memory and context management
Multi-agent	Connect, delegate to, and hand off to other agents and to humans	Multi-agent orchestration
Observability	A cost row per interaction, a health snapshot, and a full record you can drill into	Observability and debugging
Safety	Your policy controls, sandboxed code, spend limits, and isolation between agents and users	Safety and isolation
Evaluation	Score answers with a judge, verify goals, run test scenarios, and compare models	Evaluation and quality

Consistent across every channel

A WhatsApp message, a scheduled run, and an API request all behave the same way: your agent assembles its context, uses its tools, produces an answer, and records the cost. The controls you set — response limits, effort, memory, safety policy — apply everywhere, so you configure your agent once and it behaves consistently no matter how a user reaches it.

Where to start

• Deciding whether to put production traffic on your agent? Read Durability and recovery and Tool execution and reliability.
• Debugging a live agent? Start at Observability and debugging.
• Building multi-agent systems? Start at Multi-agent orchestration and the peer patterns guide.