Agent

The agent is the harness’s own assistant: you send it a prompt and it works - it calls a model, uses the tools it is allowed to use, and keeps the whole conversation in memory, so you can pick it up later or search it.

The page in the app: Agent.

Try it

pcx agent "summarize what changed in this workspace"

In the app, every agent of the host (the agent, the chief, a team’s member) has one chat: its conversations, the live steps of a turn (thinking, tool calls, results), the usage. A message sent while the agent works goes in as the next turn; Stop cancels the turn and keeps what already ran.

Models

The agent calls models through the models module. Its settings have a section for each provider - OpenRouter, Anthropic, Z.AI and OpenAI - with its API key (a secret) and, optionally, the URL of its API.

• A provider is on when its key is set. Every model its API lists can be named; nothing is turned on first.
• The app’s Models page, and every setting that names a model, offer all of them, each with its provider.
• When two providers list the same model, the first in the settings’ order answers.
• A key the provider refuses is reported on that provider’s field.

The agent’s model setting is the model of its turns (empty: glm-5.3). Prices appear when the provider’s API gives them.

What an agent may use

The tools an agent has are the tools of the modules its uses list grants - the operator decides what the agent may do, and a tool outside the turn’s grant is refused, not offered. The transcript, the tool calls and the results all stay in the session’s message stream.

When a model call fails

Failure	What the agent does
rate limited, overloaded	retries, up to 3 times, with backoff
context too long	compacts the transcript once, then retries
unauthorized, quota used up, invalid, unavailable	ends the turn with the failure
the same tool failing over and over	stops the turn as repeating (third same failure in a row)

A long conversation is compacted automatically: a summary replaces the oldest turns, the tool results are cut to what matters, and a tool result over 50 KB is trimmed for the model.

Long-running turns

• A turn streams its events (started, text, tool calls, usage, the end) - pcx agent and the app both show them live.
• A turn with no data for 10 minutes is treated as lost and recovered.
• Deleting a conversation deletes the chat, not what the agent learned: the records its turns wrote stay in memory.