AgentSkills Overview

AgentSkills are portable skill folders — a SKILL.md with instructions plus optional scripts, references, and assets. AgentSkillStepDescription runs one as a chain step.

from mmar_carl import AgentSkillStepDescription, AgentSkillStepConfig, AgentSkillExecutionMode

AgentSkillStepDescription(
    number=1,
    title="Extract PDF text",
    config=AgentSkillStepConfig(
        skill="github://anthropics/skills/skills/pdf@main",
        task="Extract the text from {pdf_path} and summarise it.",
        execution_mode=AgentSkillExecutionMode.LLM_AGENT,
        input_mapping={"pdf_path": "$memory.input.pdf_path"},
    ),
)

Skill identity

The skill field accepts a URI string (or an AgentSkillSource) — see resolvers:

Form	Example
GitHub tarball	github://anthropics/skills/skills/pdf@main
Local path	/path/to/skill
Python package	module://my_pkg.skills.pdf
Skill name	pdf (searched in ~/.agents/skills/, ./.claude/skills/, …)

Execution modes

Mode	Behaviour
LLM	SKILL.md as a system prompt; a single LLM call (default).
SCRIPT	Run a bundled script directly; no LLM call.
HYBRID	Script first, LLM fallback.
SUBAGENT	Script collects data, LLM synthesises.
LLM_AGENT	Iterative tool-calling loop: the LLM calls run_script / read_file / write_file / list_resources until it has a final answer; workspace-isolated.

Key config fields

Field	Default	Purpose
skill	—	Skill URI / source.
task	—	The task text (supports {placeholders} from input_mapping).
execution_mode	LLM	One of the modes above.
input_mapping	{}	Inputs (also staged as files in LLM_AGENT mode).
output_memory_key	None	Where to store the result.
llm_max_iterations	—	Max tool-call rounds in LLM_AGENT mode.
output_capture	—	"stdout" / "files" / "both".
filter_security_terms	True	Strip password/encrypt sections from the LLM prompt.
extra_pip	[]	Packages installed before running scripts.

See sandboxing for runtime, trust_policy, and the LLM_AGENT workspace.

See also

• Skill resolvers · Sandboxing & LLM_AGENT
• AgentSkill example (PDF → analysis → PPTX).