[FEATURE] Add support for skills

[juanrh]

Problem Statement

I would like strands to support Agent Skills as described on https://claude.com/blog/skills .

Proposed Solution

No response

Use Case

That would enable loading of relevant knowledge files depending on the specific task, utilizing the context window more effectively. This would be a generic model customization tool, that would allow reusing an agent for different tasks without code changes, just adding or removing skills that are basically a bunch of markdown files with prompts

Alternatives Solutions

No response

Additional Context

No response

[juanrh]

An option could be implementing this as a tool, as in https://github.com/xiehust/strands_demos/tree/main/strands_skills_demo, and contributing this to http://github.com/strands-agents/tools

[zastrowm]

The tool approach is probably the fastest way to get something like this integrated in Strands;

Long-term we probably need to think about how to support something like this more generally, but that will involve a design most likely

[dirien]

@juanrh is there a downside when using Strands agent tools for this in terms of token usage?

[juanrh]

The tool description should consume some of the context, but it should be worthwhile because skills are a way to promote loading data into the context only when necessary. However, I'm not part of the strands team; I'm just a user, and they should know the pros and cons better than me.

[youtuyy]

Hi team, is there any update on supporting this natively in Strands? It would be beneficial to have this in agent framework.

Reference:

1. https://blog.langchain.com/using-skills-with-deep-agents/
2. https://the-decoder.com/openai-quietly-adopts-anthropics-modular-skills-framework-to-boost-agent-capabilities/

[cliren]

Native Skills Support with Abstracted Code Interpreter

We need native Agent Skills support in Strands to unlock the specification's core benefits: token efficiency through progressive disclosure (~100 tokens for all skills vs ~100+ per traditional tool) and on-demand SKILL.md loading via code execution abstraction (local code tool or AWS AgentCore CodeInterpreter tool).

Proposed approach:

agent = Agent(
    model="claude-sonnet-4",
    skills=[
        Skill(type="local", skill_id="pdf", path="./skills/pdf"),
        Skill(type="anthropic", skill_id="pptx", version="latest")
    ],
    code_interpreter="python_repl"  # or "aws_agentcore", "anthropic"
)

This enables filesystem-based skills with pluggable code interpreter backends (local, cloud, sandboxed), combining token efficiency with flexible execution environments.

[theagenticguy]

Native Skills Support with Abstracted Code Interpreter

We need native Agent Skills support in Strands to unlock the specification's core benefits: token efficiency through progressive disclosure (~100 tokens for all skills vs ~100+ per traditional tool) and on-demand SKILL.md loading via code execution abstraction (local code tool or AWS AgentCore CodeInterpreter tool).

Proposed approach:

agent = Agent(
    model="claude-sonnet-4",
    skills=[
        Skill(type="local", skill_id="pdf", path="./skills/pdf"),
        Skill(type="anthropic", skill_id="pptx", version="latest")
    ],
    code_interpreter="python_repl"  # or "aws_agentcore", "anthropic"
)

This enables filesystem-based skills with pluggable code interpreter backends (local, cloud, sandboxed), combining token efficiency with flexible execution environments.

Why is code interpreter a dependency?

[cliren]

Agent skills may reference more instructions, scripts (ex: read_sheet.py) or write code dynamically that needs be executed using a code interpreter tool. These are required to produce more deterministic outcomes. Both static scripts and dynamic code needs to run securely in an isolated code interpreter sandbox. We can keep the current approach to provide python_repl or alternative as a tool in the tools array or introduce a SkillsConfig that specifies code interpreter.

[hi-space]

I’ve shared a sample code demonstrating how to utilize Agent Skills using the tools and multi-agent patterns of Strands Agents. You might find this useful as a reference or a temporary workaround until the official skills feature is released. Hope this helps!

• https://github.com/aws-samples/sample-strands-agents-agentskills

[adhajar]

Here’s my take on the problem.

It’s genuinely hard to predict which of these approaches will end up sticking or becoming dominant by 2026–2027. That said, there’s clearly a strong push toward solving the “how-to” problem. In that context, Strands would benefit a lot from being able to automatically discover skills from a skills directory (for example, ./github/skills//SKILL.md) as a sensible default.

A directory-based approach feels especially compelling. It would allow for seamless skill additions for agents running locally on customer machines—referencing what vercel is trying to do via https://skills.sh/ and simply running npx skills add ….

I don’t think Strands should get into the business of tightly hooking into Claude-X or Kiro-X. It’s better for the platform to remain agnostic and treat those integrations as generic skills living under the same skills directory.

I also really like the proposal of having skills defined or extended by the agent configuration itself. (listed here)

Here s my take on it:

from strands import Agent
from strands_amazon_nova import NovaAPIModel
from strands_skills import SkillsDictionary, loadSkillsFromPath

model = NovaAPIModel(
   api_key=os.env("NOVA_API_KEY"),
   model_id="nova-2-lite-v1"
)

SKILLS_PATH = os.env("SKILLS_PATH")
skills: SkillsDictionary = loadSkillsFromPath(path=SKILLS_PATH)

agent = Agent(
    model=model,
    skills=skills
)

To me, this feels cleaner and more intuitive than LangChain’s approach.