How to Intermediate · 3 min read

How to build a multi-step agent with Anthropic API

Q: How to build a multi-step agent with Anthropic API

Use the anthropic.Anthropic client to create a multi-turn conversation by managing the messages array and calling client.messages.create() repeatedly with updated context. Chain multiple steps by appending user and assistant messages to simulate an agent workflow.

Quick answer

Use the anthropic.Anthropic client to create a multi-turn conversation by managing the messages array and calling client.messages.create() repeatedly with updated context. Chain multiple steps by appending user and assistant messages to simulate an agent workflow.

PREREQUISITES

Python 3.8+
Anthropic API key
pip install anthropic>=0.20

Setup

Install the Anthropic Python SDK and set your API key as an environment variable.

bash

pip install anthropic>=0.20

Step by step

This example demonstrates a multi-step agent that first summarizes text, then answers a question about it, using the claude-3-5-sonnet-20241022 model.

python

import os
import anthropic

client = anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])

system_prompt = "You are a helpful assistant."

# Step 1: Summarize the input text
messages = [
    {"role": "user", "content": "Summarize the following text:\n\nAnthropic is an AI safety and research company."}
]
response1 = client.messages.create(
    model="claude-3-5-sonnet-20241022",
    max_tokens=200,
    system=system_prompt,
    messages=messages
)
summary = response1.content[0].text
print("Summary:", summary)

# Step 2: Ask a follow-up question using the summary
messages.append({"role": "assistant", "content": summary})
messages.append({"role": "user", "content": "What is Anthropic's main focus?"})

response2 = client.messages.create(
    model="claude-3-5-sonnet-20241022",
    max_tokens=200,
    system=system_prompt,
    messages=messages
)
answer = response2.content[0].text
print("Answer:", answer)

output

Summary: Anthropic is a company focused on AI safety and research.
Answer: Anthropic's main focus is AI safety and research.

Common variations

Use async calls with client.messages.acreate() for concurrency.
Switch models, e.g., claude-3-opus-20240229 for faster responses.
Implement streaming by setting stream=True and handling partial outputs.
Manage longer conversations by truncating or summarizing earlier messages.

Troubleshooting

If you get rate limit errors, add retries with exponential backoff.
For context length errors, reduce max_tokens or shorten conversation history.
If responses are off-topic, refine the system prompt for clearer instructions.

✅

Key Takeaways

Build multi-step agents by managing the messages list to maintain conversation context.
Use the claude-3-5-sonnet-20241022 model for best coding and reasoning performance.
Append assistant and user messages between calls to simulate multi-turn workflows.
Handle errors by managing token limits and refining system prompts.
Async and streaming modes enable more efficient and interactive agents.

Verified 2026-04 · claude-3-5-sonnet-20241022

Verify ↗