Agent Abilities by Anthropic on Oct 16, 2025, as a approach to prolong Claude with reusable capabilities. Inside months, the idea gained traction throughout the AI group and started evolving right into a broader design sample for constructing modular, transportable agent capabilities past Claude itself.
As an AI practitioner, I’ve been utilizing Claude Code for fairly a while and have seen many tutorials explaining create Agent Abilities inside the Claude ecosystem. Nonetheless, once I tried to implement the identical idea for our enterprise answer with out counting on Claude, I rapidly bumped into a unique set of design questions. What precisely defines an “agent ability”? How ought to it’s structured? And the way ought to abilities be triggered and orchestrated in a customized agent framework?
In Claude, a typical Agent Ability is outlined as a Markdown file containing a reputation, description, directions, and scripts. This natural-language interface works nicely in lots of conditions, particularly when ambiguity is appropriate. However in manufacturing techniques, precision usually issues greater than flexibility. One recurring problem I encountered was ability triggering: typically a ability can be highly effective and well-designed, but Claude would fail to invoke it as a result of the outline was too imprecise.
For our brokers, I would like stricter ensures—abilities that set off deterministically and behave persistently each time. That requirement led me to implement abilities straight in Python with express logic. However doing so raised an attention-grabbing architectural query: if a ability is applied purely in code, how is it completely different from a software, a characteristic, or simply one other perform? Even when these abilities are reusable and invoked on demand, what truly makes them agent abilities?
This text explores these questions and shares a sensible perspective on designing and implementing agent abilities for customized AI brokers with out counting on Claude.
Agent abilities vs Instruments vs Options
A software is a primitive functionality, or a single motion the agent can take. Every software does one particular factor. Instruments are particular person devices, like hammers, saws, and drills.
- Instance: Claude’s instruments embrace issues like
bash_tool(run a command),web_search(search the web),view(learn a file),str_replace(edit a file),web_fetch(get a webpage),places_search,weather_fetchand so forth.
A ability is a set of directions for orchestrate a number of instruments to perform a fancy process nicely. A ability doesn’t give me new capabilities. It provides brokers experience in combining present instruments successfully. Abilities are like an in depth recipe that tells the brokers which devices to make use of, in what order, and what errors to keep away from.
- Instance: The docx ability, for instance, tells Claude to make use of
bash_toolto runnpm set up docx, then write JavaScript utilizing particular patterns, then runbash_toolonce more to validate the output, then usepresent_filesto share it.
A characteristic is a product-level idea, or one thing the consumer sees and may toggle on or off. A characteristic is enabled by giving the agent entry to sure instruments and abilities.
- Examples: “Code Execution and File Creation,” “Internet Search,” and “Artifacts” are options. So “file creation” as a characteristic is powered by the bash software, the file creation software, varied doc abilities, and the present_files software, all working collectively.
Abilities are what bridge the hole between uncooked software entry and high-quality output. With out the docx ability, Claude may nonetheless technically create a Phrase doc, however it would possibly miss issues like “all the time set web page measurement explicitly as a result of docx-js defaults to A4” or “by no means use unicode bullets.”
Do agent abilities should be within the format of markdown recordsdata?
Not essentially. The idea of agent abilities is broader than the format.
A ability is essentially codified experience for accomplish a process nicely. That experience may dwell in lots of varieties:
- A markdown file with directions (typical Claude Agent Ability)
- A Python script that does the work straight
- A config file or JSON schema
- A set of instance inputs and outputs
- A mix of all the above
In my case, the Python script appears applicable as a result of I would like deterministic and dependable execution each single time with no variation. It’s sooner, cheaper, and extra predictable for a procedural course of. The markdown instruction method turns into useful when the duty includes ambiguity or judgment. Typically, we want LLM studying directions and reasoning about what to do subsequent so as to add worth over a inflexible script.
A hybrid method can be widespread. I can hold my Python code and downgrade it to a software implementation, however add a markdown ability file that helps the brokers perceive when to invoke my agent ability and interpret the outcomes. A typical product iteration would possibly start with a Python implementation and step by step incorporate Markdown directions, finally forming a hybrid ability design.
Agent ability vs MCP? Agent Ability + MCP!
Our brokers already connect with information sources by way of MCP servers, however a lot of our agent abilities additionally embrace the flexibility to learn straight from databases. This raises a sensible architectural query: when ought to an agent use MCP, and when ought to the aptitude dwell inside a ability?
Anthropic explains the excellence clearly with a useful kitchen analogy:
MCP connects Claude to information; Abilities teaches Claude what to do with that information.
- MCP gives the skilled kitchen – entry to instruments, substances, and gear.
- Abilities present the recipes – directions that inform the agent use these sources to provide one thing useful.
In my design, I deal with MCP connections as infrastructure and abilities as orchestration on how the information is used.
MCP servers are answerable for exposing exterior information sources and companies. For instance, they might present structured entry to databases, logs, APIs, or inner techniques. Their position is to make these sources obtainable to the agent in a standardized means.
Agent abilities, alternatively, outline how the agent ought to use the information from MCP servers and different databases to perform a process.
For that reason, I sometimes implement:
- Database entry, APIs, and information retrieval as instruments (or MCP capabilities)
- Choice logic and workflows as agent abilities
Abilities as “Agentic RAG”
Ideally, the agent abilities load info dynamically from instruments when they’re executed. This makes the sample similar to agentic retrieval-augmented era (RAG).
As a substitute of preloading all context into the immediate, the ability can:
- Establish what info it wants
- Retrieve the related information by way of a software or MCP server
- Course of that information in keeping with its directions
- Produce the ultimate output
This method retains the brokers light-weight whereas nonetheless permitting abilities to entry giant or altering datasets on demand.
Conclusion
Anthropic launched an vital paradigm shift, and Claude’s implementation of Agent Abilities gives useful inspiration for constructing our personal agent techniques. So long as our abilities seize the core function of what a ability is supposed to do, which is encapsulating reusable on-demand capabilities for an agent, the particular format and implementation particulars can differ. Finally, the design choices round when and use abilities ought to be guided by the wants and constraints of the product we’re constructing.
Thanks for studying! I hope this has been useful to you.
