Plugins and other agents

plugins and agents are at the top of the file, before the first [section]. [card] is a section.

plugins = ["plugins/acme-support"]
agents = [
  { id = "billing", url = "https://billing-agent.example.com" },
]

[temporal]
# ...

plugins

The plugin folders to load. A plugin gives the agent tools, skills, prompts, hooks and commands. Add tools, commands and skills explains how to make one.

• Default: none. The agent has no tools.
• Example: plugins = ["plugins/acme-support", "/opt/shared/plugins/crm"]
• Relative paths start from the folder where you start tiny-harness.
• If a plugin is not valid, tiny-harness does not load it. If one part of a plugin is not valid, tiny-harness skips that part and loads the rest.

agents

Other A2A agents that this agent can give work to. The agent sends them a sub-task and waits for the result.

• Default: none.
• tiny-harness reads the agent card of each agent at start. It refuses an agent that asks for authentication or for a protocol extension that tiny-harness does not know.
• The agent sees a list of these agents. It decides itself when to give work to one. What the agent sees explains the list.

Each agent has these settings:

Setting	Default	What it does
id	none: you must set it	The name the agent uses for this other agent
url	none: you must set it	The address of the other agent
version	none	The version that tiny-harness records for the other agent
updates	"auto"	How tiny-harness learns about changes in a waiting sub-task
follow_interval	"PT5S" (5 seconds)	With follow: the pause before tiny-harness listens again. At least 1 second

What the agent sees

An agent card is a file that another agent publishes about itself. tiny-harness reads these values from each card and shows them to the agent:

• the id from your configuration file
• the name and the description from the card
• the name and the description of each skill on the card, up to 10 skills

The agent uses the description and the skills to decide when to give work to the other agent. Thus, a clear description helps the agent choose correctly.

If the other agent is also tiny-harness, write its description in its own [card] section.

tiny-harness does not show the url to the agent. It treats the card text as data, not as instructions. It removes line breaks and the characters < and >. It cuts a long value and adds …. A name can have 80 characters, a description 500 and a skill description 200. When tiny-harness cuts a value, it writes a warning to the log at start.

The list also has the agent itself, with the id self. The agent gives work to itself when a step needs a clean context: for example, research or a long tool result that the main task does not need to keep. The sub-task gets only its goal, description and acceptance criteria, and it sends back its answer. You cannot use self as the id of an agent in this file.

The agent can give work only to an agent in this list. If it names a different agent, the sub-task does not start and the agent gets an error.

updates

A sub-task on another agent can wait for input. While it waits, it can change: for example, a person answers it on the other agent. This setting chooses how tiny-harness learns about the change.

Value	What happens	Use it when
"auto"	The other agent sends updates to this server. If the other agent cannot send updates, tiny-harness uses "follow"	The other agent can reach this server's base_url
"follow"	tiny-harness keeps a connection open to the other agent and listens	The other agent cannot reach this server
"off"	tiny-harness learns about a change only at its next message	You do not need updates between messages

agents = [
  { id = "billing", url = "https://billing.example.com", updates = "follow", follow_interval = "PT10S" },
]

[card]

How other agents see this agent. tiny-harness publishes an agent card at /.well-known/agent-card.json. Another agent that lists this one under agents reads the card and shows its name, description and skills to its model.

• Default: the name tiny-harness, a general description and no skills. Another agent's model then cannot know what this agent is for.
• Set a description that says what this agent does and what it needs. Add a skill for each kind of work that it does.

[card]
name = "Order ledger"
description = "Answers questions about the status of an order. Give it the order number."

[[card.skills]]
id = "order-status"
name = "Order status"
description = "Looks up an order by its number and reports its status and dates."
examples = ["What is the status of order 48213?"]

Setting	Default	What it does
name	"tiny-harness"	The name on the card
description	A general description of tiny-harness	What this agent does
version	"0.1.0"	The version on the card
skills	none	The skills on the card. Each skill has an id and a name, and can have a description, tags and examples

A value that you set must not be empty. Two skills must not have the same id. If they do, tiny-harness does not start, and it names the setting.

Example: two agents shows two tiny-harness agents. One gives work to the other.