Skip to content

Product

The agent harness

An agent is a model plus a harness. The model reasons; Haarness gives it everything else — durable state, tools, context, permissions and a record of what it did — in one runtime that every client shares.

haarness
01

Durable state

A versioned server owns every thread, turn and tool call. Clients come and go; turns, child agents and workflows keep running, and a restart never replays a tool.

02

Tools and delegation

Read, search, edit, patch, shell and web tools, media-aware reads, durable child agents, dynamic workflows, persistent goals and native schedules.

03

Context and skills

AGENTS.md, Agent Skills, slash commands, custom agents, memory and compaction keep long chats grounded in your project.

04

Permissions and budgets

Ordered allow, ask and deny rules for every tool, blocking approvals and questions, and execution budgets you set. Policy you can read, not a black box.

05

Any model

Native OpenAI, Anthropic, Gemini, Bedrock, Copilot, Vertex AI and Azure adapters, any OpenAI-compatible endpoint, or the managed router.

06

Observability

Exportable transcripts, cross-project stats, a private insights report and OpenTelemetry traces of every turn, request and tool call.