Tools and skills

Turning off built-in tools

litellm.agent(Harness.CLAUDE_CODE, task, sandbox=box, model="litellm_proxy/coder", disable_tools=["web_search", "bash"])

Use the normalized names from Events. Codex can't filter its built-in tools, so this raises CapabilityUnsupported there.

Custom tools

Custom Python tools are supported on Deep Agents in this release. They're plain functions that run in your process.

from typing import Literal

def lookup_owner(path: str) -> str:
    """Return the CODEOWNERS entry for a file path."""
    return codeowners.match(path)

async def open_ticket(title: str, body: str, severity: Literal["low", "high"]) -> str:
    """File a ticket in the tracker. Returns the ticket URL."""
    return await tracker.create(title=title, body=body, severity=severity)

litellm.agent(
    Harness.DEEPAGENTS,
    "Triage the failing test and file a ticket.",
    tools=[lookup_owner, open_ticket],
    sandbox=box,
    model="litellm_proxy/coder",
)

The type hints and docstring become the tool's schema. Passing tools= to Claude Code, Codex or OpenCode raises CapabilityUnsupported.

Skills

A skill is a local folder containing a SKILL.md and any files it needs. Pass folder paths, and the adapter copies each one into the sandbox where the runtime looks for skills.

litellm.agent(Harness.CLAUDE_CODE, "Ship the fix.", sandbox=box, model="litellm_proxy/coder", skills=["./skills/release-checklist"])

Harness	Installed to
CLAUDE_CODE	.claude/skills/<name>/ in the working directory
CODEX	$CODEX_HOME/skills/<name>/
OPENCODE	.opencode/skill/<name>/ in the working directory
DEEPAGENTS	passed to the agent

Skills are copied once when the session starts.