建议agent 支持skill

[xcpuma]

skill 已经成为agent 能力的必备，建议直接支持

[Maplemx]

收到，已经在做了

[Maplemx]

Progress Report:

1. We'd tried to append document reading abilities and build SkillsExecutor component base on reading abilities. But we find that can only fit pure text skills. If we want to execute skills with references including MCP server, bash command or dependency package installations, that'll beyond the SkillsExecutor's handle ability.

2. So now we're trying to refactor more core modules like ToolManager. We want to separate action commands and actual action executions to ensure future action ability extendable. We're still working on it.

[Maplemx]

Progress Report:

After v4.1.1, we've refactored core module ToolManage(now it's Action) and TriggerFlow to get ready for SkillsExecutor.

The early version of SkillsExecutor has been developed and it works as we expected. But still, there're many other works to be done like testing basic features, designing nice API methods for better development experience, writing and testing real world examples.

Hope to present this new SkillsExecutor feature soon.

[Maplemx]

Progress Report:

from v4.1.2 -> v4.1.2.5, we improved many important features including TriggerFlow core refactor, Action core refactor, Runtime Environment, DAG Planner and Executor and finally we achieved Skills Executor!

You can use skills as example:
https://github.com/AgentEra/Agently/blob/main/examples/skills_executor/06_executable_education_course_pack.py

Or you can just provide actions(MCP, local function) and skills then let the agent decide how to work as example:
https://github.com/AgentEra/Agently/blob/main/examples/agent_auto_orchestration/18_amap_mcp_trip_planner.py

We're redesign the outer methods and APIs to improve the development experience as the last step for v4.1.3.

Hope to present v4.1.3 to you soon.

[Maplemx]

Good news, v4.1.3 has come with brand new Skills Executor and Agent Auto Orchestration!

Full release story: https://github.com/AgentEra/Agently/blob/main/docs/en/development/release-notes-4.1.3.md

agent.start() is now the default user-layer entrypoint for a candidate-aware Agent turn. The caller still receives the business result, while the Agent can route through ordinary model response, Actions, Skills Executor, or Dynamic Task when those candidates were explicitly declared.

result = (
    agent
    .use_actions([lookup_customer, fetch_contract, notify_owner])
    .use_skills(
        [{"source": "anthropics/skills", "subpath": "skills/docx"}],
        mode="model_decision",
    )
    .use_dynamic_task(mode="auto", max_tasks=8)
    .input({"customer_id": "C-1024", "ticket": "payment failure"})
    .output({
        "summary": (str, "business summary", True),
        "risk_level": (str, "low / medium / high", True),
        "next_actions": ([str], "recommended actions", True),
    })
    .start()
)