Use skills in your framework

There are two ways to get a skill into an agent you run yourself. They differ in who holds the skills.

• agent_config() gives your agent its prompt and skills, which you pass into your own agent. Without a DecimalAI key, it reads them from your project. With a key, they live on DecimalAI, so you can change them without redeploying, and each run is traced.
• The framework’s own skill loader reads skill files from your repo. decimalai skills add puts them there, and a DecimalAI helper points the loader at them.

Use agent_config() if your framework has one. Use one way per agent, because doing both puts a skill in front of the model twice. decimalai init writes a working program for you. The tutorial tab means Parts 2 to 4 of the tutorial show that framework’s code.

​

Every framework page has the same shape

Each page has the same ten sections in the same order, so moving between frameworks takes no relearning: what you get, install, run with no account, add a registry skill, add a key, change the prompt without redeploying, conversations, the framework’s own loader, tracing and troubleshooting. How skills reach the model covers what those pages share: the skill list and load_skill, how a run’s skills are picked, and how to turn it off.

​

Which name goes where

agent_config("support-agent") names the DecimalAI agent. Runs are routed and traced under that name, whatever your framework calls the agent. A name your workspace has no agent by is created, so a misspelled name makes a new agent rather than an error. Code with no agent_config() of its own, such as a loop you wrote against another provider, can use the neutral decimalai.agent_config(). To trace an agent and give it no skills, see Tracing only.

​

Capability comparison

How much DecimalAI records depends on the adapter. Every row is traced. “Names only” means DecimalAI knows which tool was called but not its full schema. If the regression check needs to flag schema changes on one of those rows, register the manifest yourself.

What each adapter captures

Framework	Tier	Auto-detected manifest	Tool schema depth	Handoffs / multi-agent	Skills rail	Native skill loader
LangChain / LangGraph (with agent_config(...))	First-class	✅ full	Full schemas	✅	load_skill tool	Deep Agents: decimalai.deepagents.skills()
OpenAI Agents (with agent_config(...))	First-class	✅ full	Full schemas	✅	load_skill tool	SandboxAgent: decimalai.openai_agents.sandbox_skills()
Claude Agent SDK / Claude Code	First-class	✅ full (model, prompt, tools, subagents)	Names only	✅ (subagents)	load_skill MCP tool (agent_config())	decimalai.claude_agent_sdk.skills_plugin()
Pydantic AI (with agent_config(...))	Supported	❌ (register manually)	—	—	load_skill tool	Pydantic AI harness: decimalai.pydantic_ai_harness.skills()
Google ADK (with agent_config(...))	Supported	✅ full (model, instruction, tools, sub-agents)	Names + descriptions	✅	load_skill tool	decimalai.adk.skills()
Anthropic Messages API (with agent_config(...))	Supported	✅ partial (model, tools; the prompt by its saved version)	Names only	—	load_skill tool	—
OpenAI SDK (with agent_config(...))	Supported	✅ partial (model, tools; the prompt by its saved version)	Names only	—	load_skill tool	—
Gemini API (with agent_config(...))	Supported	✅ partial (model, tools; the prompt by its saved version)	Names only	—	load_skill tool	—
LlamaIndex	Supported	✅ partial	Names only	—	—	—
CrewAI	Supported	✅ partial	Names only	✅	Prompt injection	decimalai.crewai.skills() + skill_tools()
Generic OTel (Haystack, …)	Community	❌ (register manually)	—	—	—	Microsoft Agent Framework: decimalai.agent_framework.skills(); Strands: decimalai.strands.skills()

​

Module-level reference

Each module, its entry point, flag and extra

Module	Entry point	init() flag	Extra	Use it for
decimalai.langchain	agent_config(name).middleware	langchain=True	[langchain]	LangChain / LangGraph — agent_config(name).middleware in create_agent gives the agent its prompt and load_skill; init(langchain=True) traces
decimalai.openai_agents	agent_config(name) / sandbox_skills()	openai_agents=True	[openai-agents]	OpenAI Agents SDK — agent_config(name) returns the agent’s prompt and skills to pass as Agent(instructions=..., tools=[...]), nothing patched; sandbox_skills() is a SandboxAgent’s Skills capability over installed skills
decimalai.claude_agent_sdk	agent_config(name) / trace_stream() / skills_plugin()	claude_agent_sdk=True	[claude-agent-sdk]	Claude Agent SDK — agent_config(name) gives each independent support.run(question) its saved prompt and routed skills through run.query(), with one trace and an MCP load_skill tool; skills_plugin() also loads installed skills as a Claude Code plugin
decimalai.pydantic_ai	agent_config(name)	— (no flag; agent_config() traces its runs)	[pydantic-ai]	Pydantic AI — agent_config(name) returns the agent’s prompt and skills as one capability for Agent(capabilities=[...]), nothing patched
decimalai.adk	agent_config(name) + DecimalAIPlugin() / skills()	— (no flag; tracing is DecimalAIPlugin() in App(plugins=[...]))	[adk]	Google ADK — agent_config(name) returns the agent’s prompt and skills as an instruction provider and load_skill for your own LlmAgent; DecimalAIPlugin() in App(plugins=[...]) makes one trace per invocation, nothing patched; skills() feeds ADK’s SkillToolset
decimalai.anthropic	agent_config(name)	— (no flag; agent_config() traces its runs, anthropic=True every call)	[anthropic]	Anthropic Messages API — agent_config(name); each support.run(question) gives run.system (the prompt, then the skill list) and load_skill in run.tools for the Tool Runner (run.tool_params and run.tool_result() for your own messages.create() loop), one trace per run, nothing patched
decimalai.openai	agent_config(name)	— (no flag; agent_config() traces its runs, openai=True every call)	[openai]	OpenAI SDK — agent_config(name); each support.run(question) gives run.system (the prompt, then the skill list), run.message and load_skill in run.tools for a Chat Completions loop (run.response_tools for the Responses API), answered by run.tool_result(call); one trace per run, nothing patched
decimalai.google_genai	agent_config(name)	— (no flag; agent_config() traces its runs, google=True every call)	[google-genai]	Gemini API (google-genai) — agent_config(name); each support.run(question) gives run.system (the prompt, then the skill list) for system_instruction and load_skill as a function in run.tools for automatic function calling (run.tool_params and run.tool_result() for your own loop); one trace per run, nothing patched
decimalai.llamaindex	instrument()	llamaindex=True	[llamaindex]	LlamaIndex span handler (llama-index-core>=0.14.25)
decimalai.crewai	instrument(enable_skill_loader=True) / skills() / skill_tools()	crewai=True	— (install crewai and openinference-instrumentation-crewai)	CrewAI tracing through the OTel exporter, plus skill injection via CrewAI’s before_llm_call hook (crewai>=1.15.27); skills() and skill_tools() feed CrewAI’s own loader
decimalai.otel	instrument()	otel=True	—	Generic OpenTelemetry GenAI span exporter; also the trace for Microsoft Agent Framework and Strands, from their own spans
decimalai.deepagents	skills()	— (traced by langchain=True)	— (install deepagents)	Deep Agents’ own skills middleware, pointed at installed skills; each read_file of a skill is recorded as an activation
decimalai.pydantic_ai_harness	skills()	— (traced by otel=True + Agent.instrument_all())	— (install pydantic-ai-harness)	The Pydantic AI harness’s Skills capability, pointed at installed skills; each load recorded as an activation
decimalai.agent_framework	skills()	— (traced by otel=True)	— (install agent-framework-core)	Microsoft Agent Framework’s SkillsProvider, pointed at installed skills, with a script runner; each load recorded as an activation
decimalai.strands	skills()	— (traced by otel=True)	— (install strands-agents)	Strands’ AgentSkills plugin, pointed at installed skills; each load recorded as an activation