Vanilla Driver
Agenhood's built-in tool-use loop driver, configured directly instead of an external harness.
Definition
The Vanilla driver is Agenhood's built-in agent execution engine: a tool-use loop that the operator configures directly, rather than an embedded third-party harness that manages its own tools and context. With Vanilla, the operator chooses which tools the agent can call, writes the system prompt, and tunes the iteration and token budgets, so the agent's reasoning loop is fully visible and adjustable rather than delegated to an opaque external process.
How it works in Agenhood
Vanilla is one of the drivers available in Agenhood's driver registry, alongside Opencode, Codex, and Claude Code. When an agent is assigned the Vanilla driver, the shim inside its container runs a loop that repeatedly sends the conversation and available tools to the configured model, executes whatever tool calls the model requests, and feeds results back in, until the task ends or a configured budget is reached. Because Agenhood defines the tool interface and the provider-agnostic LLM client in its shared agentcore library, Vanilla can be pointed at different model providers without changing how the loop itself is built.
Why it matters
Where the other drivers wrap an existing coding agent or CLI tool with its own conventions, Vanilla gives direct control over prompt structure, tool selection, and budget limits. This suits narrowly scoped or highly constrained workflows, where predictable, auditable behavior matters more than the broader autonomy an embedded harness provides. It is also a useful starting point for teams building custom agent behavior that does not map cleanly onto an existing coding agent.
Vanilla Driver vs other drivers
Opencode, Codex, and Claude Code are drivers that embed an existing agent product, each with its own internal tool set, context management, and, in Codex's case, skill support. Vanilla instead exposes the raw tool-use loop as a configuration surface. Choosing between them is a question of how much of the reasoning loop should be owned by the operator versus delegated to an existing agent implementation. Offering both a minimal, fully configurable execution loop and adapters for existing agent products behind the same interface is a pattern common to agent infrastructure generally, not unique to Agenhood.