Agents and chat

Agents

Agents are persistent AI team members with a role, instructions, model assignment, department, tools, and operating state. Venture agents serve one venture; organization-level agents are managed separately.

You can:

• generate an initial team from venture context;
• create an agent manually;
• edit its name, role, instructions, department, model, and tools;
• activate or pause an agent;
• inspect agent details and activity;
• remove agents when authorized.

Agent changes affect future work. They do not rewrite completed workflow history.

Model and tool configuration

Model choices come from organization AI settings and the configured provider credentials. Tools determine which platform capabilities an agent can invoke. Grant only the tools required by the role.

If a selected model or credential becomes unavailable, work may fail even though the agent remains active. Review Settings, then retry the work.

Tool visibility combines venture/workflow context, department support, agent role, capabilities, guardrails, and approved resources. Naming a tool in an agent prompt does not grant access to it.

See Agent tool access (MCP) for the permission model and Compute, tokens, and model selection for provider/model precedence.

Chat

Venture chat lets users communicate with the AI team in venture context. Conversations and messages are persisted, and the interface can stream responses as they are generated.

Chat can use venture knowledge and agent context. It is best for questions, analysis, and guided actions; use Pipelines for structured, auditable multi-step work.

Authorized leadership chat sessions can request approved resource access. Credentials are supplied programmatically rather than through chat text; see Resources and Security Officer.

What to expect

• Chat is for working with the AI team, not human-to-human messaging.
• Tool access is configured through agent capabilities and guardrails.
• Organization-level agent controls can require a different role from venture agent controls.