Skip to content
AdiosDocumentation
Browse documentation

AgentKit

AgentKit describes an agent's instructions, inputs, outputs, workflows, functions, and connections as files. Use the Agent Builder to configure and run an agent, or edit the same files with your development tools.

The portability goal is one agent definition with the same contracts locally and on Adios. Three choices belong to the run environment:

ChoiceOptions
Agent executionYour machine or Adios hosted
Function executionYour machine, Adios Sandbox, or Adios hosted
CredentialsExplicit local credentials or authorized Adios connections

Adios Sandbox is the local environment for testing execution closer to Adios hosting. Running an agent locally does not imply that its functions or connected services must also run locally.

Start here

What travels with the kit

agentkit.yaml names the agents, entrypoints, workflows, and contracts. Referenced files hold instructions, JSON schemas, function source, and test fixtures. Installation settings, connection bindings, credentials, and run history belong to the environment. Keep secret values out of kit files and model instructions.

A workflow coordinates steps. An agent step calls a model; a function executes code; a connector accesses a service. Artifact steps save files, and approval steps wait for a decision. These are separate responsibilities even when one machine runs them all.

Current support

Local and hosted agent workflows share the workflow engine and structured model response contract. Native function tests and a separate hosted function runtime also exist. Hosted Agent Builder workflows currently reject function steps; automatic selection of local or hosted function execution is still being connected.

Local model credentials can be supplied through a named environment variable. Saved agents can use Adios connections during local runs. General secret pulling and an AgentKit executor for Adios Sandbox are not yet exposed as one complete workflow. The guides mark these boundaries so examples remain runnable.

Equal behavior means matching schemas, dependencies, approvals, artifacts, retries, and effect handling. Different model providers can produce different text; use recorded responses when testing the workflow itself.