Agent Skills: Create an Atomic Agent

Create an Atomic Agent

An agent is an LLM-backed transformer from one BaseIOSchema to another. Building one means: design the schemas, write the system prompt, wire the provider client, build the AgentConfig, instantiate AtomicAgent[In, Out].

For deep material (streaming, token counting, hooks, multi-agent memory), the authority is ../framework/references/agents.md plus providers.md, prompts.md, and memory.md. This skill is the action-oriented path: clarify → write → run.

When this fires vs the umbrella framework skill

• This skill: the user is creating or wiring a specific agent — "add a planner agent", "build a Q&A agent", "make a router that classifies tickets".
• framework skill: questions about Atomic Agents in general, or the user is doing something other than authoring an agent.

Phase 1 — Clarify

Bundle into one message:

1. What should the agent do? One sentence. Becomes the persona / background line.
2. Inputs and outputs. Use BasicChatInputSchema / BasicChatOutputSchema for free-form chat. Use a custom pair for anything structured (extraction, classification, planning, routing). When custom, branch to the create-atomic-schema skill for the schema authoring.
3. Provider. OpenAI / Anthropic / Groq / Ollama / Gemini / OpenRouter / MiniMax. Default: whatever the project already uses; otherwise OpenAI.
4. Conversational? Yes → wire a ChatHistory. No (single-shot transformer) → omit it for stateless behavior.
5. Context providers. Anything to inject into the prompt at runtime (current time, user identity, retrieved docs)? If yes, plan to also use the create-atomic-context-provider skill afterwards.

Skip anything already settled in context.

Phase 2 — Plan

State the plan in one short block:

• File: <project>/agents/<agent_name>.py (or directly in main.py for a tiny project — see ../framework/references/project-structure.md).
• Schemas: which pair, where they live.
• Provider + model + Instructor mode. Default models: OpenAI gpt-5-mini, Anthropic claude-haiku-4-5, Groq llama-3.3-70b-versatile, Ollama llama3.1, Gemini gemini-2.5-flash.
• SystemPromptGenerator content — three sections: background, steps, output_instructions.
• History? Hooks? Context providers?

Phase 3 — Implement

Canonical imports (do not deviate)

from atomic_agents import (
    AtomicAgent, AgentConfig,
    BasicChatInputSchema, BasicChatOutputSchema,
)
from atomic_agents.context import ChatHistory, SystemPromptGenerator
from instructor import Mode

Wire the provider client (always Instructor-wrapped)

The full per-provider matrix lives in ../framework/references/providers.md. Quick recap:

# OpenAI — default mode is Mode.TOOLS
import os, instructor, openai
client = instructor.from_openai(openai.OpenAI(api_key=os.environ["OPENAI_API_KEY"]))
model = "gpt-5-mini"
api_params: dict = {}

# Anthropic — Mode.TOOLS, max_tokens REQUIRED in model_api_parameters
import anthropic
client = instructor.from_anthropic(anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"]))
model = "claude-haiku-4-5"
api_params = {"max_tokens": 4096}

# Gemini — Mode.GENAI_TOOLS, assistant_role="model"
from google import genai
client = instructor.from_genai(genai.Client(api_key=os.environ["GEMINI_API_KEY"]), mode=Mode.GENAI_TOOLS)
model = "gemini-2.5-flash"
api_params = {}

# Groq / Ollama / MiniMax — Mode.JSON in both factory and AgentConfig

Build the agent

from atomic_agents import AtomicAgent, AgentConfig
from atomic_agents.context import ChatHistory, SystemPromptGenerator

agent = AtomicAgent[MyInput, MyOutput](
    config=AgentConfig(
        client=client,
        model=model,
        history=ChatHistory(),  # omit for stateless
        system_prompt_generator=SystemPromptGenerator(
            background=["You are a concise research assistant."],
            steps=[
                "Read the question carefully.",
                "Decide what minimum information answers it.",
                "Produce the answer in the required schema.",
            ],
            output_instructions=[
                "Reply under 100 words.",
                "If unsure, set status='error' and explain why.",
            ],
        ),
        # Provider-specific knobs — match the Instructor factory
        # mode=Mode.TOOLS,                         # OpenAI / Anthropic / OpenRouter
        # mode=Mode.JSON,                          # Groq / Ollama / MiniMax
        # mode=Mode.GENAI_TOOLS, assistant_role="model",  # Gemini
        model_api_parameters=api_params or {"temperature": 0.2},
    )
)

Generics carry the truth

AtomicAgent[MyInput, MyOutput] — write the type parameters explicitly. The framework reads them at class-definition time. Do not rely on subclass-level input_schema / output_schema class attributes.

Provider-specific knobs (most common gotchas)

• Anthropic without max_tokens in model_api_parameters → API rejects every call.
• Gemini without assistant_role="model" → role mismatch on every turn.
• Groq / Ollama / MiniMax with Mode.TOOLS → tools formatted in a way the provider does not accept; flip to Mode.JSON.
• Reasoning models (o-series, GPT-5 reasoning variants) → often want system_role=None and reasoning_effort in model_api_parameters.

Phase 4 — Run and verify

out = agent.run(MyInput(...))
print(out)

Quick smoke test without paying for a real call:

uv run python -c "from <project>.agents.<agent_name> import agent; print(type(agent).__name__, '->', agent.input_schema.__name__, '/', agent.output_schema.__name__)"

If output validation fails repeatedly, the parse:error hook has the details — see ../framework/references/hooks.md for registration.

Phase 5 — Hand off

Tell the user:

• How to call agent.run(...) (and run_async, run_stream, run_async_stream when appropriate).
• Which env var to set for the provider key.
• Optional next steps:
  • Tools the agent should be able to invoke → create-atomic-tool skill.
  • Dynamic data injected into the prompt → create-atomic-context-provider skill.
  • Custom schemas → create-atomic-schema skill.
  • Multiple agents working together → ../framework/references/orchestration.md.
  • Telemetry / retries / logging → ../framework/references/hooks.md.
  • Conversation persistence, summarization, multi-agent memory → ../framework/references/memory.md.

Anti-patterns

• Forgetting to wrap the client with instructor.from_* — structured outputs silently stop working.
• BaseModel instead of BaseIOSchema for the agent's input or output type.
• AgentConfig.mode out of sync with the Instructor factory mode.
• assistant_role="assistant" on Gemini — must be "model".
• Missing max_tokens on Anthropic — every call fails.
• Hardcoded API keys in the source — read from env.
• Unbounded ChatHistory in a long-running service — monitor agent.get_context_token_count().utilization or set max_messages.

For deep material — streaming, async, token counting, hooks, multi-agent history — load ../framework/references/agents.md.