What happens when a conversation starts
- xpander picks the harness: the one the message names, else the conversation’s last one, else the agent’s default. A new agent’s default follows the organization’s default model: Anthropic (directly or through Amazon Bedrock) gives Claude Code, OpenAI gives Codex, anything else gives OpenCode.
- An executor takes the turn. Turns in one conversation run one at a time. Conversations of the same agent run in parallel.
- The turn runs in an isolated sandbox. It can’t see other agents’ files or processes. The agent’s directory is at
/agent/data, and the conversation works insessions/<conversation id>. - The harness reaches its model through xpander, with a short-lived token. It never gets a provider key. xpander applies the model policy and budget and records the call.
- xpander writes the instructions and connects the skills, then streams the turn into the conversation. Actions that need sign-off appear as approval cards.
Instructions and memory
Each turn’s prompt combines the agent’s instructions, its skills, its notes, the person’s memories and the task context, up to 256 KiB. When it is too long, task context and skill descriptions are trimmed first, never the instructions or the output contract. Claude Code reads it through--append-system-prompt-file. Codex and OpenCode read it as AGENTS.md. The agent keeps long-term notes in CLAUDE.md and AGENTS.md at the top of its directory. Tell it to remember something and it writes there.
How skills reach the session
- API connectors and custom actions are served over MCP as
mcp__xpander__<action>. xpander attaches the credential, applies approvals and records each call. The session never holds the credential. - Remote MCP connectors use the same endpoint. Local MCP connectors are added to the harness’s own configuration.
- Markdown skills sync into
.claude/skills(Claude Code) or~/.agents/skills(Codex and OpenCode). The agent’s own skills live inlocal_skills/. - Workspace secrets arrive as environment variables, refreshed before every turn.
What is isolated, what is shared
Each conversation has its own working directory and harness session. Each turn runs in a sandbox that holds this agent’s directory only, so a conversation can open the folders of the agent’s other conversations. All conversations of an agent share its installed packages, skills, notes and theshared/ folder. Two agents never share a directory.

The agent's directory in the Computer view: sessions/ (one folder per conversation), shared/ and skills/ next to the agent's own AGENTS.md and CLAUDE.md. output/ is a folder this agent made for itself; xpander does not create it. Shown with sample data.
Between messages
The harness stays running between your messages while someone has the conversation open, so a follow-up reaches the same process with no start-up: Claude Code on its open input, Codex through its app server and OpenCode throughopencode serve. OpenCode follow-ups also reuse the prompt cache, because a conversation keeps the same temporary folder from turn to turn. After two minutes with nobody in the conversation, the process parks: it saves its session and frees its capacity. Your next message wakes it in the same session and files.
When a session ends
A turn ends when the harness finishes. A sub-task another agent started also ends after 15 minutes with no output. Your next message resumes the session. A spent budget stops the run with “Execution stopped: budget exceeded.” Closing the tab does not stop a turn. Stop ends what the agent is doing now and keeps the session open.
Confirming Stop on a running turn.
Next steps
Create and customize an agent
Instructions, harness, model and skills.
Work with tasks and files
The task record and results.
Decide who can use an agent
Permissions on agents and conversations.

