Skills

Skills are named, versioned, sandboxed execution units — think of them as the agent's procedural memory. Where tools are individual operations, skills package multi-step workflows, tool combinations, and conditional logic into a callable unit that can be discovered, invoked, and tracked for performance.

Overview

Skills are higher-level, sandboxed execution units compared to tools. While tools are primitive operations, skills can be multi-step procedures with LLM integration, custom instructions, and streaming execution events.

Protocol Messages (5 types)

Type String	Model	Direction
skill/discover/req	SkillDiscoverRequest	Agent → Host
skill/discover/resp	SkillDiscoverResponse	Host → Agent
skill/call/req	SkillCallRequest	Agent → Host
skill/call/resp	SkillCallResponse	Host → Agent
skill/event	SkillEvent	Host → Agent (streaming)

SkillDefinition

Comprehensive skill manifest:

Field	Type	Description
name	str	Unique skill name
version	str	Skill version
description	str	What the skill does
triggers	list[str]	When to activate this skill
tools	list[str]	Required tools
toolkits	list[str]	Required toolkits
status	SkillStatus	Created, Blocked, Published, Archived
input_schema	dict	JSON Schema for input
output_schema	dict	JSON Schema for output
instructions	str	LLM instructions/prompt
file_path	str	Skill file location
llm_config	LLMConfig	LLM provider settings
arguments	dict	Default arguments
when_to_use	str	Usage hint
argument_hint	str	Argument guidance
source	str	Origin
category	str	Skill category
tags	list[str]	Search tags
max_turns	int	Max execution turns
timeout_seconds	float	Execution timeout
icon	str	Display icon
metadata	dict	Extra metadata

SkillResult

Field	Type	Description
success	bool	Execution succeeded
data	dict	Result data
summary	str	Human-readable summary
truncated	bool	Output truncated
error	str	Error message
error_code	SkillErrorCode	UNKNOWN_SKILL, SKILL_ERROR, RUNTIME_ERROR
duration_ms	int	Execution time
events	list[SkillEvent]	Streaming events collected

SkillEvent (extends A2EEvent)

Streaming events during skill execution with kinds: skill.started, tool.started, tool.completed, tool.failed, llm.started, llm.token, llm.completed, skill.completed, skill.failed.

LLMConfig

Field	Type	Description
provider_name	str	LLM provider (e.g. "anthropic")
provider_credentials	dict	API keys
provider_config	dict	Provider-specific settings
is_default	bool	Use as default LLM

SkillPlugin ABC

class SkillPlugin(A2EPlugin):
    @abstractmethod
    def _list_skills(self) -> list[SkillDefinition]: ...

    @abstractmethod
    def _execute_skill(self, name, arguments, context) -> SkillResult: ...

    def discover(self, msg):
        # List skills then filter by tags and categories

    def call(self, msg):
        # Execute skill with streaming support
        # Creates inner emit_event() closure
        # Passes context dict: {emit_event, llm_override, metadata, streaming}

SkillAPI (Client)

from a2e.caps.skills.client import SkillAPI

skills = SkillAPI(client)

# Discover available skills
skill_list = skills.discover(tags=["analysis"], categories=["data"])

# Call a skill
result = skills.call(
    name="example-analysis",
    arguments={"data": "sample"},
    streaming=True,
    on_event=lambda event: print(f"Event: {event.kind}"),
    timeout=60.0
)

if result.success:
    print(result.data)