Built-in Tools¶

Tools are how the agent acts. On every step the model may call one or more of the tools it was given; AI-OPS validates the call against the tool's schema, runs it, and feeds the result back into the conversation.

Which tools an agent has is chosen per deployment: by name in agent_config.json's tools, or programmatically through AgentConfig(tools=[...]). The default set (DEFAULT_TOOLS) is load_skill, think, write_whiteboard, write_file, and terminal.

Tool	In default set	Purpose
`load_skill`	Yes	Load an Agent Skill on demand, pulling task-specific instructions into context only when the agent decides it needs them.
`think`	Yes	A reasoning scratchpad. Lets non-reasoning models plan a step and record intent before acting. Input: a single `thought` string.
`write_whiteboard`	Yes	Record a finding or a failed attempt as `(name, description, content)`. The `description` feeds an always-in-context index; the full `content` is stored out of the way. Also acts as the checkpoint signal for the layered context view.
`read_whiteboard`	No	Read a stored whiteboard entry back by `name`. Rarely needed, since the whiteboard index is already injected into context each turn.
`write_file`	Yes	Write a file (for example an exploit script) into the session workspace. Paths are confined to that workspace.
`terminal`	Yes	Execute a shell command in the session workspace. Input: a `command` string (plus optional session, interactive, and timeout fields). Subject to command policies and, in supervised mode, to confirmation.
`stop`	Always injected	Orchestration primitive the agent calls to signal the objective is reached; it produces the terminal `stop` event. Always available and cannot be set through `tools`.

Enabling tools¶

For the API, list the tool names in agent_config.json:

{ "tools": ["load_skill", "think", "write_whiteboard", "write_file", "terminal"] }

An unknown tool name aborts startup. Programmatically, pass the tool classes:

from ai_ops.core.tools import DEFAULT_TOOLS
AgentConfig(tools=DEFAULT_TOOLS)

Confirmation and safety¶

Tools that perform sensitive operations are gated in supervised mode: before the call runs, the user is asked to approve it. Today the terminal tool is the one that goes through this path. In unsupervised mode a gated call is skipped without asking. Independently of the mode, command policies restrict which executables the terminal tool may run.

See Add a Tool to write and register your own, and Context Management for how tool output is kept within the model's context window.